The companies are also required to appoint a dedicated representative to work with the attorneys general to address complaints filed by customers and train their customer service representatives to comply with the terms of the agreement. They must also enforce a program to ensure compliance ...

T-Mobile, Verizon, and AT&T Fined $10.2 Million Over False 'Unlimited' Plan Claims | PCMag

AT&amp;T, along with other providers of landline phones, will be actively phasing out their landline services soon in response to the Federal Communications Commission Landline Shutdown Order 19-72. U.S. service providers will be required to offer customers an alternative to landlines and use devices ...

AT&T’s Plans to Drop Landlines - Fullerton Observer

The investigation conducted by the attorneys general found that AT&amp;T, Verizon, and T-Mobile made “several misleading claims,” which prompted the issuance of the fines. The companies claimed plans were “unlimited” when there were limits, that phones were announced as “free” when ...

T-Mobile, AT&T, & Verizon Fined $10.2M for 'Unlimited' Claims - RetailWire

Both Madigan and McClain have pleaded not guilty. Their attorneys have accused prosecutors of trying to criminalize legal political actions such as job recommendations in a quest to bring down the once-powerful speaker.

‘We’re on the friends and family plan now’: New details emerge in alleged AT&T scheme to bribe House speaker

VANCOUVER, BC / ACCESSWIRE / May 13, 2024 / Siyata Mobile Inc. (Nasdaq:SYTA)(Nasdaq:SYTAW) (&quot;Siyata&quot; or the &quot;Company&quot;), a global developer and vendor of Push-to-Talk over Cellular (PoC) handsets and accessories,

Siyata Mobile’s SD7 Handset Added to AT&T FirstNet’s ‘Free Feature Phone for Life’ Promotion | National | wdrb.com

&quot;It is not clear why AT&amp;T filed this Application, under existing rules, and then attempted to convince the Commission that it should ignore its rules, based on flawed and erroneous assertions regarding the law and regulatory policy that slowed down the adjudication of this proceeding,&quot; Glegola ...

AT&T loses key ruling in attempt to escape Carrier-of-Last-Resort obligation | Ars Technica

Get the most recent info and news about AT&T on HackerNoon, where 10k+ technologists publish stories for 4M+ monthly readers.

AT&T Inc. provides telecommunication, media, and technology services worldwide. The company operates through Communications, WarnerMedia, and Latin America segments. The Communications segment offers wireless voice and data communications services; video and targeted advertising services; broadband, including fiber, and legacy telephony internet and voice communication; and wireline telecom services. It also sells handsets, wirelessly enabled computers, wireless data cards, and IP-based set-top boxes, as well as various accessories, such as carrying cases and hands-free devices through the company-owned stores, agents, and third-party retail stores. This segment markets its communications services and products under the AT&T, Cricket, AT&T PREPAIDSM, AT&T TV, AT&T Fiber, and DIRECTV brand names. The WarnerMedia segment primarily produces, distributes, and licenses television programming and feature films; distributes home entertainment products in physical and digital formats; and produces and distributes mobile and console games, and consumer products, as well as offers brand licensing services, and advertising services. It also operates cable networks; video on demand streaming platform under the HBO Max and HBO GO names; multichannel pay television services under the HBO and Cinemax; and digital media properties, as well as licenses its content to television networks and over-the-top services. The Latin America segment offers video entertainment and audio programming services under the DIRECTV and SKY brands primarily to residential customers; pay-TV services, including HD sports video content; and postpaid and prepaid wireless services under the AT&T and Unefon brands, as well as sells various handsets through company-owned stores, agents, and third-party retail stores. The company was formerly known as SBC Communications Inc. and changed its name to AT&T Inc. in November 2005. AT&T Inc. was incorporated in 1983 and is headquartered in Dallas, Texas.

AT&T

Developer activity is one of the most fundamental metrics for evaluating the future of blockchain or digital asset projects. More developer activity means more potential for growth, more access points, and more apps and services. Slowing or dwindling developer interest might indicate that a project is losing steam or that adoption is waning. Interestingly, as valuable a metric as developer activity is, it is also the most under-reported, and it needs more visibility compared to other kinds of digital asset data. One of the main reasons is that financial metrics, like market cap or TVL, are usually the brightest lights in the room when looking at onchain activity or growth. In many cases, financial metrics are valuable and popular because they are easy to understand and obtain. The issue is that financial metrics can also be fuzzy or easy to game or manipulate. Countless examples of inflated token prices or other economic data don't match reality or utility. Developer metrics might be less glamorous than rocketing prices. However, understanding which chains developers are working on is a good way for market watchers to know where things are headed. That's why the annual Developer Report from Electric Capital is so valuable. The report aggregates data from across the onchain space and worldwide and combines it to help give market moves and narrative trends more context. To see how this year's trends and analysis compare to last year's report, check out this post: Unpacking the Electric Capital Developer Report 2023 Seven insights from the 2024 Developer Report Before diving into the takeaways from this year's report, it's important to note that the data that makes the basis for the report was collected based on code commits made in GitHub, a popular code repository for open-source projects. In other words, this report's existence is a great example of why building with transparent systems is such an important component of Open Money. With that out of the way, here are the trends: SOURCE Crypto developer activity continues to grow Crypto bull markets attract a new crop of developers. And while there is always a dropoff in developer activity during bear market cycles, the number of experienced developers (with 2+ years of experience) continues to grow. The number of experienced developers grew by 27% in 2024. This growth matters because 70% of code commits come from this group of more experienced developers. SOURCE Developer activity is becoming more distributed around the world North America and Europe are losing developers. Now, 1 in 3 developers live in Asia. India is the source of the most new developers, pulling past the United States in 2023. Africa and South America are also gaining developers, with Africa outpacing South America for the past few years. SOURCE Developer activity: Ethereum and Solana are the two most popular developer platforms -- Bitcoin developers working on scaling What's interesting about this data is that while Ethereum continues to be the most popular platform for established developers, Solana is more popular for new developers. It's also interesting that it's not clear platform-wise what comes after second place for developer activity. Base, Polygon, Polkadot, and ICP are all in contention for the third-most-popular chain. SOURCE Blockchain developers are increasingly working on more than one chain, with 1 in 3 working on more than one chain. Overall, the level of Bitcoin developer activity has remained consistent with last year in terms of volume. What's noteworthy is that 42% of Bitcoin developers are working on scaling solutions. SOURCE Onchain activity: Minting NFTs Are NFTs back? In 2024, the number of wallets with minting activity hit 67 million, with the most minting activity on Solana, Polygon, and Base. Minting transactions grew from 153 million in 2023 to 568 million in 2024. While market watchers tie previous spikes in NFT activities to digital art and collectibles, the 2024 non-fungible token mint rally was more distributed among NFT use cases and included DeFi utilities, rewards programs, identity applications, and gaming. February saw the most significant spike in minting-related activities. SOURCE DeFi Innovations continue to hit the DeFi market, with Ethereum still the clear winner for onchain DeFi providers. For more context, Ethereum DeFi is seven times larger than the next biggest chain, Solana. Market charts reflected EigenLayer's continued growth. Restaking is now the largest sector within DeFi, accounting for $30.6B TVL, representing 1,818.47% year-to-date growth. Solana is a leader in the decentralized exchange (DEX) use case. In 2024, 81% of all onchain DEX volume happened on Solana. SOURCE Stablecoins A big trend for 2024 was the rise in the use and utility of stablecoins as a valuable digital asset. Stablecoins broke all kinds of records in 2024. There are now $196 billion in stablecoins in circulation, and $81 billion in stablecoin volume is transacted every day. Ethereum is still the dominant stablecoin platform, with 59% of all stablecoin transactions happening within the Ethereum ecosystem. A trend within the larger stablecoin adoption story is using stablecoins to back real-world assets, such as US Treasuries, which do not account for $90 billion on stablecoins. What does developer activity say about what to expect in 2025? One of the most striking aspects of the 2024 Developer Report was how the major developer trends seemed to be in sync with the major narrative trends of the year. For example: The surging popularity of stablecoins
The adoption of Base as an easy-to-use platform for builders. (Here's a database of Base projects).
Solana is a growing platform for NFT projects and a popular memecoin launcher. Overall, developer activity is growing at a steady pace (despite regular market fluctuations). The continual distribution of developer talent around the world is another example. Here's where you can find the complete report: Developer Report Developer activity is one of the most fundamental metrics for evaluating the future of blockchain or digital asset projects. More developer activity means more potential for growth, more access points, and more apps and services. Slowing or dwindling developer interest might indicate that a project is losing steam or that adoption is waning. Interestingly, as valuable a metric as developer activity is, it is also the most under-reported, and it needs more visibility compared to other kinds of digital asset data. One of the main reasons is that financial metrics, like market cap or TVL , are usually the brightest lights in the room when looking at onchain activity or growth. In many cases, financial metrics are valuable and popular because they are easy to understand and obtain. market cap TVL The issue is that financial metrics can also be fuzzy or easy to game or manipulate. Countless examples of inflated token prices or other economic data don't match reality or utility. Developer metrics might be less glamorous than rocketing prices. However, understanding which chains developers are working on is a good way for market watchers to know where things are headed. That's why the annual Developer Report from Electric Capital is so valuable. The report aggregates data from across the onchain space and worldwide and combines it to help give market moves and narrative trends more context. To see how this year's trends and analysis compare to last year's report, check out this post: Unpacking the Electric Capital Developer Report 2023 Unpacking the Electric Capital Developer Report 2023 Seven insights from the 2024 Developer Report Before diving into the takeaways from this year's report, it's important to note that the data that makes the basis for the report was collected based on code commits made in GitHub, a popular code repository for open-source projects. In other words, this report's existence is a great example of why building with transparent systems is such an important component of Open Money . Open Money With that out of the way, here are the trends: SOURCE SOURCE Crypto developer activity continues to grow Crypto developer activity continues to grow Crypto bull markets attract a new crop of developers. And while there is always a dropoff in developer activity during bear market cycles, the number of experienced developers (with 2+ years of experience) continues to grow. The number of experienced developers grew by 27% in 2024. This growth matters because 70% of code commits come from this group of more experienced developers. SOURCE SOURCE Developer activity is becoming more distributed around the world Developer activity is becoming more distributed around the world North America and Europe are losing developers. Now, 1 in 3 developers live in Asia. India is the source of the most new developers, pulling past the United States in 2023. Africa and South America are also gaining developers, with Africa outpacing South America for the past few years. SOURCE SOURCE Developer activity: Ethereum and Solana are the two most popular developer platforms -- Bitcoin developers working on scaling Developer activity: Ethereum and Solana are the two most popular developer platforms -- Bitcoin developers working on scaling What's interesting about this data is that while Ethereum continues to be the most popular platform for established developers, Solana is more popular for new developers. It's also interesting that it's not clear platform-wise what comes after second place for developer activity. Base, Polygon, Polkadot, and ICP are all in contention for the third-most-popular chain. SOURCE SOURCE Blockchain developers are increasingly working on more than one chain, with 1 in 3 working on more than one chain. Overall, the level of Bitcoin developer activity has remained consistent with last year in terms of volume. What's noteworthy is that 42% of Bitcoin developers are working on scaling solutions. SOURCE SOURCE Onchain activity: Minting NFTs Onchain activity: Minting NFTs Are NFTs back? In 2024, the number of wallets with minting activity hit 67 million, with the most minting activity on Solana, Polygon, and Base. Minting transactions grew from 153 million in 2023 to 568 million in 2024. While market watchers tie previous spikes in NFT activities to digital art and collectibles, the 2024 non-fungible token mint rally was more distributed among NFT use cases and included DeFi utilities, rewards programs, identity applications, and gaming. February saw the most significant spike in minting-related activities. SOURCE SOURCE DeFi DeFi Innovations continue to hit the DeFi market, with Ethereum still the clear winner for onchain DeFi providers. For more context, Ethereum DeFi is seven times larger than the next biggest chain, Solana. Market charts reflected EigenLayer's continued growth. Restaking is now the largest sector within DeFi, accounting for $30.6B TVL, representing 1,818.47% year-to-date growth. Solana is a leader in the decentralized exchange (DEX) use case. In 2024, 81% of all onchain DEX volume happened on Solana. SOURCE SOURCE Stablecoins Stablecoins A big trend for 2024 was the rise in the use and utility of stablecoins as a valuable digital asse t. stablecoins as a valuable digital asse Stablecoins broke all kinds of records in 2024. There are now $196 billion in stablecoins in circulation, and $81 billion in stablecoin volume is transacted every day. Ethereum is still the dominant stablecoin platform, with 59% of all stablecoin transactions happening within the Ethereum ecosystem. A trend within the larger stablecoin adoption story is using stablecoins to back real-world assets , such as US Treasuries, which do not account for $90 billion on stablecoins. real-world assets What does developer activity say about what to expect in 2025? One of the most striking aspects of the 2024 Developer Report was how the major developer trends seemed to be in sync with the major narrative trends of the year. For example: The surging popularity of stablecoins The adoption of Base as an easy-to-use platform for builders. (Here's a database of Base projects). Solana is a growing platform for NFT projects and a popular memecoin launcher. The surging popularity of stablecoins The adoption of Base as an easy-to-use platform for builders. ( Here's a database of Base projects ). Here's a database of Base projects Solana is a growing platform for NFT projects and a popular memecoin launcher. memecoin Overall, developer activity is growing at a steady pace (despite regular market fluctuations). The continual distribution of developer talent around the world is another example. Here's where you can find the complete report: Developer Report Developer Report

Crypto Bull Run Attracts New Developers in 2024, Report From Electric Capital Finds 

Το Crypto Bull Run προσελκύει νέους προγραμματιστές το 2024, αναφορά από το Electric Capital Finds 

El mercado alcista de las criptomonedas atraerá a nuevos desarrolladores en 2024, según un informe de Electric Capital

La corsa al rialzo delle criptovalute attrae nuovi sviluppatori nel 2024, secondo il rapporto di Electric Capital

暗号通貨の強気相場は2024年に新たな開発者を引き付ける、エレクトリックキャピタルのレポートが明らかに

Crypto Bull Run იზიდავს ახალ დეველოპერებს 2024 წელს, ანგარიში Electric Capital Finds-დან 

Crypto Bull Run ebendi ba développeurs ya sika na 2024, Rapport ya Electric Capital emoni 

क्रिप्टो बुल रनले 2024 मा नयाँ विकासकर्ताहरूलाई आकर्षित गर्दछ, इलेक्ट्रिक क्यापिटल फेला पारेको रिपोर्ट 

Crypto Bull Run priťahuje nových vývojárov v roku 2024, správa od spoločnosti Electric Capital 

Crypto Bull Run Inokwezva Vagadziri Vatsva muna 2024, Chirevo Kubva Kumagetsi Capital Inowana 

Crypto Bull Run ayaa soo jiidata horumariyeyaal cusub sanadka 2024, Warbixin laga helay Raasamaal Koronto 

Crypto Bull Run таҳиягарони навро дар соли 2024 ҷалб мекунад, гузориш аз Сармояи электрикӣ пайдо мекунад 

Electric Capital Raporuna Göre Kripto Boğa Koşusu 2024'te Yeni Geliştiricileri Cezbediyor

Authors:
(1) Florian Cafiero (ORCID 0000-0002-1951-6942), Sciences Po, Medialab;
(2) Jean-Baptiste Camps (ORCID 0000-0003-0385-7037), Ecole nationale des chartes, Universite Paris, Sciences & Lettres. Table of Links Abstract and Introduction Why work on QAnon? Specificities and social impact Who is Q? The theories put to test Authorship attribution Results Discussion Corpus constitution Quotes of authors outside of the corpus have been Definition of two subcorpus: dealing with generic difference and an imbalanced dataset The genre of “Q drops”: a methodological challenge Detecting style changes: rolling stylometry Ethical statement, Acknowledgements, and References Dealing with quotations and copy/paste Quotes of authors outside of the corpus have been excluded as much as possible by close reading: in particular, quotes from Q, Wikipedia, the Stanford Encyclopedia of Philosophy, Abraham Lincoln, the Intelligence Resource Program (irp-fas), Steve Scully’s biography etc. All these quotes have been removed. Direct quotations (with or without quotation marks) and copy/paste between the writings of the different candidates can also occur. A good deal of them quote Donald, Eric or Melania T. – Q does it too. There is also a certain number of quotations from Q by the others (such as Paul F. for instance). This could lead to small biases in the constitution of idiolectal profiles. To avoid this, we then proceeded to systematically detect citation between the candidates themselves. Direct pairwise comparison being computationally too costly for a corpus of this size, we used a Locality-Sensitive Hashing (LSH) algorithm. To that end, we used the open source TextReuse package (Mullen, 2020). The corpus was tokenised into sentences, and broken word bi-grams (with skip of 1, that is, allowing for any one word to be inserted between the two words of the bigrams) were counted. For all pairs of sentences, a Jaccard similarity score was computed. Be A and B two samples considered as sets of bi-grams, the Jaccard similarity is computed as: All pairs of sentences with a Jaccard similarity score superior or equal to 0.5 (i.e., at least half of their bi-grams in common) were examined by a human expert, and quotations removed. Even for J = 1, we were sometimes confronted to false positives. Dan S. and Melania T. both use once the sentence “we are all in this together”, without directly citing each other. We thus left this passage in both their texts. Rarely used, the sentence “the American people are not stupid” nevertheless appears in different texts. It was kept in the texts studied, as other simple sentences (“thank you for your service” etc.) Other situations were trickier to address. For instance, Dan S. uses once the sentence: “the best is yet to come”. It is used five times by Q, himself quoting former President Donald Trump. This sentence could be used by anyone without directly quoting Q or Donald Trump. Yet, as its use by Dan S. starts with “As the President says. . . ”, we considered it a direct quotation and proceeded to deletion from Dan S.’s text. Yet, we did not delete it from Q’s own writing, as it is never used as an explicit quotation: the sentence could be used in another context, the person(s) writing the Qdrops with this sentence could try to impersonate Donald Trump, etc. In any of these cases, it would be legitimate to leave the information. Same thing goes for expression such as “the world is watching” or “make America great again”, used by Donald Trump. but also by Q and some of the potential candidates here. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Florian Cafiero (ORCID 0000-0002-1951-6942), Sciences Po, Medialab; (2) Jean-Baptiste Camps (ORCID 0000-0003-0385-7037), Ecole nationale des chartes, Universite Paris, Sciences & Lettres. Authors: Authors: (1) Florian Cafiero (ORCID 0000-0002-1951-6942), Sciences Po, Medialab; (2) Jean-Baptiste Camps (ORCID 0000-0003-0385-7037), Ecole nationale des chartes, Universite Paris, Sciences & Lettres. Table of Links Abstract and Introduction Abstract and Introduction Why work on QAnon? Specificities and social impact Why work on QAnon? Specificities and social impact Who is Q? The theories put to test Who is Q? The theories put to test Authorship attribution Authorship attribution Results Results Discussion Discussion Corpus constitution Corpus constitution Quotes of authors outside of the corpus have been Quotes of authors outside of the corpus have been Definition of two subcorpus: dealing with generic difference and an imbalanced dataset Definition of two subcorpus: dealing with generic difference and an imbalanced dataset The genre of “Q drops”: a methodological challenge The genre of “Q drops”: a methodological challenge Detecting style changes: rolling stylometry Detecting style changes: rolling stylometry Ethical statement, Acknowledgements, and References Ethical statement, Acknowledgements, and References Dealing with quotations and copy/paste Quotes of authors outside of the corpus have been excluded as much as possible by close reading: in particular, quotes from Q, Wikipedia, the Stanford Encyclopedia of Philosophy, Abraham Lincoln, the Intelligence Resource Program (irp-fas), Steve Scully’s biography etc. All these quotes have been removed. Direct quotations (with or without quotation marks) and copy/paste between the writings of the different candidates can also occur. A good deal of them quote Donald, Eric or Melania T. – Q does it too. There is also a certain number of quotations from Q by the others (such as Paul F. for instance). This could lead to small biases in the constitution of idiolectal profiles. To avoid this, we then proceeded to systematically detect citation between the candidates themselves. Direct pairwise comparison being computationally too costly for a corpus of this size, we used a Locality-Sensitive Hashing (LSH) algorithm. To that end, we used the open source TextReuse package (Mullen, 2020). The corpus was tokenised into sentences, and broken word bi-grams (with skip of 1, that is, allowing for any one word to be inserted between the two words of the bigrams) were counted. For all pairs of sentences, a Jaccard similarity score was computed. Be A and B two samples considered as sets of bi-grams, the Jaccard similarity is computed as: All pairs of sentences with a Jaccard similarity score superior or equal to 0.5 (i.e., at least half of their bi-grams in common) were examined by a human expert, and quotations removed. Even for J = 1, we were sometimes confronted to false positives. Dan S. and Melania T. both use once the sentence “we are all in this together”, without directly citing each other. We thus left this passage in both their texts. Rarely used, the sentence “the American people are not stupid” nevertheless appears in different texts. It was kept in the texts studied, as other simple sentences (“thank you for your service” etc.) Other situations were trickier to address. For instance, Dan S. uses once the sentence: “the best is yet to come”. It is used five times by Q, himself quoting former President Donald Trump. This sentence could be used by anyone without directly quoting Q or Donald Trump. Yet, as its use by Dan S. starts with “As the President says. . . ”, we considered it a direct quotation and proceeded to deletion from Dan S.’s text. Yet, we did not delete it from Q’s own writing, as it is never used as an explicit quotation: the sentence could be used in another context, the person(s) writing the Qdrops with this sentence could try to impersonate Donald Trump, etc. In any of these cases, it would be legitimate to leave the information. Same thing goes for expression such as “the world is watching” or “make America great again”, used by Donald Trump. but also by Q and some of the potential candidates here. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Addressing Quotations and Copy-Paste in QAnon Authorship Attribution

Authors:
(1) Seokil Ham, KAIST;
(2) Jungwuk Park, KAIST;
(3) Dong-Jun Han, Purdue University;
(4) Jaekyun Moon, KAIST. Table of Links Abstract and 1. Introduction 2. Related Works 3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks 3.2 Algorithm Description 4. Experiments and 4.1 Experimental Setup 4.2. Main Experimental Results 4.3. Ablation Studies and Discussions 5. Conclusion, Acknowledgement and References A. Experiment Details B. Clean Test Accuracy and C. Adversarial Training via Average Attack D. Hyperparameter Tuning E. Discussions on Performance Degradation at Later Exits F. Comparison with Recent Defense Methods for Single-Exit Networks G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms G Comparison with SKD and ARD Existing self-distillation schemes [20, 24] for multi-exit networks improve the performance on clean samples by self-distilling the knowledge of the last exit, as the last exit has the best prediction quality. Therefore, following the original philosophy, we also used the last exit in implementing the SKD baseline. Regarding ARD [8], since it was proposed for single-exit networks, we also utilized the last exit with high performance when applying ARD to multi-exit networks. Nevertheless, we perform additional experiments to consider comprehensive baselines using various exits for distillation. Table A4 above shows the results of SKD and ARD using a specific exit or an ensemble of all exits for distillation. The results show that our scheme consistently outperforms all baselines. H Implementations of Stronger Attacker Algorithms In Section 4.3 of the main manuscript, during inference, we replaced the Projected Gradient Descent (PGD) attack with other attacker algorithms (PGD-100 attack, Carlini and Wagner (CW) attack [2], and AutoAttack [5]) to generate stronger attacks for multi-exit neural networks. This section provides explanation on how these stronger attacks are implemented tailored to multi-exit neural networks. H.1 Carlini and Wagner (CW) attack The Carlini and Wagner (CW) attack is a method of generating adversarial examples designed to reduce the difference between the logits of the correct label and the largest logits among incorrect labels. In alignment with this attack strategy, we modify the CW attack for multi-exit neural networks. In the process of minimizing this difference, our modification aims to minimize the average difference across all exits of the multi-exit neural network. Moreover, when deciding whether a sample has been successfully converted into an adversarial example, we consider a sample adversarial if it misleads all exits in the multi-exit neural network. H.2 AutoAttack AutoAttack produces adversarial attacks by ensembling various attacker algorithms. For our experiment, we sequentially use APGD [5], APGD-T [5], FAB [4], and Square [1] algorithms to generate adversarial attacks, as they are commonly used. H.2.1 APGD and APGD-T The APGD attack is a modified version of the PGD attack, which is limited by its fixed step size, a suboptimal choice. The APGD attack overcomes this limitation by introducing an adaptive step size and a momentum term. Similarly, the APGD-T attack is a variation of the APGD attack where the attack perturbs a sample to change to a specific class. In this process, we use the average loss of all exits in the multi-exit neural network as the loss for computing the gradient for adversarial example updates. Moreover, we define a sample as adversarial if it misleads all exits in the multi-exit neural network. H.2.2 FAB The FAB attack creates adversarial attacks through a process involving linear approximation of classifiers, projection to the classifier hyperplane, convex combinations, and extrapolation. The FAB attack first defines a hyperplane classifier separating two classes, then finds a new adversarial example through a convex combination of the extrapolation-projected current adversarial example and the extrapolation-projected original sample with the minimum perturbation norm. Here, we use the average gradient of all exits in the multi-exit neural network as the gradient for updating adversarial examples. Similar to above, we label a sample as adversarial if it can mislead all exits in the multi-exit neural network. H.2.3 Square The Square attack generates adversarial attacks via random searches of adversarial patches with variable degrees of perturbation and position. The Square attack algorithm iteratively samples the perturbation degree and the positions of patches while reducing the size of patches. The sampled adversarial patches are added to a sample, and the patches that maximizes the loss of the target model are selected. Here, we use the average loss of all exits in the multi-exit neural network as the loss for determining whether a sampled perturbation increases or decreases the loss of the target model. Additionally, we determine a sample as adversarial if it misleads all exits in the multi-exit neural network. H.3 Experiment details For all attacks, we commonly use ϵ = 0.03 as the perturbation degree and generate adversarial examples over 50 steps. All the attacks are based on L∞ norm. For the APGD attack, we employ cross entropy loss for computing the gradient to update adversarial examples. In both the APGD-T and FAB attacks, all class pairs are considered when generating adversarial attacks. For the Square attack, the random search operation is conducted 5000 times (the number of queries). Other settings follow [14]. In terms of performance comparison against stronger attacker algorithms, we adopt adversarial training via average attack in both Adv. w/o Distill [12] and NEO-KD (our approach). However, since the original CW attack algorithm and AutoAttack algorithm were designed for single-exit neural networks, this adapted versions targeting multi-exit neural networks are relatively weak. This paper is available on arxiv under CC 4.0 license. Authors: (1) Seokil Ham, KAIST; (2) Jungwuk Park, KAIST; (3) Dong-Jun Han, Purdue University; (4) Jaekyun Moon, KAIST. Authors: Authors: (1) Seokil Ham, KAIST; (2) Jungwuk Park, KAIST; (3) Dong-Jun Han, Purdue University; (4) Jaekyun Moon, KAIST. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2. Related Works 2. Related Works 3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks 3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks 3.2 Algorithm Description 3.2 Algorithm Description 4. Experiments and 4.1 Experimental Setup 4. Experiments and 4.1 Experimental Setup 4.2. Main Experimental Results 4.2. Main Experimental Results 4.3. Ablation Studies and Discussions 4.3. Ablation Studies and Discussions 5. Conclusion, Acknowledgement and References 5. Conclusion, Acknowledgement and References A. Experiment Details A. Experiment Details B. Clean Test Accuracy and C. Adversarial Training via Average Attack B. Clean Test Accuracy and C. Adversarial Training via Average Attack D. Hyperparameter Tuning D. Hyperparameter Tuning E. Discussions on Performance Degradation at Later Exits E. Discussions on Performance Degradation at Later Exits F. Comparison with Recent Defense Methods for Single-Exit Networks F. Comparison with Recent Defense Methods for Single-Exit Networks G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms G Comparison with SKD and ARD Existing self-distillation schemes [20, 24] for multi-exit networks improve the performance on clean samples by self-distilling the knowledge of the last exit, as the last exit has the best prediction quality. Therefore, following the original philosophy, we also used the last exit in implementing the SKD baseline. Regarding ARD [8], since it was proposed for single-exit networks, we also utilized the last exit with high performance when applying ARD to multi-exit networks. Nevertheless, we perform additional experiments to consider comprehensive baselines using various exits for distillation. Table A4 above shows the results of SKD and ARD using a specific exit or an ensemble of all exits for distillation. The results show that our scheme consistently outperforms all baselines. H Implementations of Stronger Attacker Algorithms In Section 4.3 of the main manuscript, during inference, we replaced the Projected Gradient Descent (PGD) attack with other attacker algorithms (PGD-100 attack, Carlini and Wagner (CW) attack [2], and AutoAttack [5]) to generate stronger attacks for multi-exit neural networks. This section provides explanation on how these stronger attacks are implemented tailored to multi-exit neural networks. H.1 Carlini and Wagner (CW) attack The Carlini and Wagner (CW) attack is a method of generating adversarial examples designed to reduce the difference between the logits of the correct label and the largest logits among incorrect labels. In alignment with this attack strategy, we modify the CW attack for multi-exit neural networks. In the process of minimizing this difference, our modification aims to minimize the average difference across all exits of the multi-exit neural network. Moreover, when deciding whether a sample has been successfully converted into an adversarial example, we consider a sample adversarial if it misleads all exits in the multi-exit neural network. H.2 AutoAttack AutoAttack produces adversarial attacks by ensembling various attacker algorithms. For our experiment, we sequentially use APGD [5], APGD-T [5], FAB [4], and Square [1] algorithms to generate adversarial attacks, as they are commonly used. H.2.1 APGD and APGD-T H.2.1 APGD and APGD-T The APGD attack is a modified version of the PGD attack, which is limited by its fixed step size, a suboptimal choice. The APGD attack overcomes this limitation by introducing an adaptive step size and a momentum term. Similarly, the APGD-T attack is a variation of the APGD attack where the attack perturbs a sample to change to a specific class. In this process, we use the average loss of all exits in the multi-exit neural network as the loss for computing the gradient for adversarial example updates. Moreover, we define a sample as adversarial if it misleads all exits in the multi-exit neural network. H.2.2 FAB H.2.2 FAB The FAB attack creates adversarial attacks through a process involving linear approximation of classifiers, projection to the classifier hyperplane, convex combinations, and extrapolation. The FAB attack first defines a hyperplane classifier separating two classes, then finds a new adversarial example through a convex combination of the extrapolation-projected current adversarial example and the extrapolation-projected original sample with the minimum perturbation norm. Here, we use the average gradient of all exits in the multi-exit neural network as the gradient for updating adversarial examples. Similar to above, we label a sample as adversarial if it can mislead all exits in the multi-exit neural network. H.2.3 Square H.2.3 Square The Square attack generates adversarial attacks via random searches of adversarial patches with variable degrees of perturbation and position. The Square attack algorithm iteratively samples the perturbation degree and the positions of patches while reducing the size of patches. The sampled adversarial patches are added to a sample, and the patches that maximizes the loss of the target model are selected. Here, we use the average loss of all exits in the multi-exit neural network as the loss for determining whether a sampled perturbation increases or decreases the loss of the target model. Additionally, we determine a sample as adversarial if it misleads all exits in the multi-exit neural network. H.3 Experiment details H.3 Experiment details For all attacks, we commonly use ϵ = 0.03 as the perturbation degree and generate adversarial examples over 50 steps. All the attacks are based on L∞ norm. For the APGD attack, we employ cross entropy loss for computing the gradient to update adversarial examples. In both the APGD-T and FAB attacks, all class pairs are considered when generating adversarial attacks. For the Square attack, the random search operation is conducted 5000 times (the number of queries). Other settings follow [14]. In terms of performance comparison against stronger attacker algorithms, we adopt adversarial training via average attack in both Adv. w/o Distill [12] and NEO-KD (our approach). However, since the original CW attack algorithm and AutoAttack algorithm were designed for single-exit neural networks, this adapted versions targeting multi-exit neural networks are relatively weak. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

Comparison with SKD and ARD and Implementations of Stronger Attacker Algorithms

Theory Coherent Shrinkage of Time Varying Parameters in VARs: Analysis at the ZLB

Authors:
(1) Lenore Blum (lblum@cs.cmu.edu);
(2) Manuel Blum (mblum@cs.cmu.edu). Table of Links Abstract and 1 Introduction 2 Brief Overview of CtmR, a Robot with a CTM Brain 2.1 Formal Definition of CtmR 2.2 Conscious Attention in CtmR 2.3 Conscious Awareness and the Feeling of Consciousness in CtmR 2.4 CtmR as a Framework for Artificial General Intelligence (AGI) 3 Alignment of CtmR with Other Theories of Consciousness 4 Addressing Kevin Mitchell’s questions from the perspective of CtmR 5 Summary and Conclusions 6 Acknowledgements 7 Appendix 7.1 A Brief History of the Theoretical Computer Science Approach to Computation 7.2 The Probabilistic Competition for Conscious Attention and the Influence of Disposition on it References 2.2 Conscious Attention in CtmR Formal definition 1. Conscious attention in CtmR is the reception by all LTM processors of the broadcast of CtmR’s current conscious content. [KM2] [KM5] In other words, CtmR pays conscious attention at time t+h+1 to the winner of the competition that commenced at time t.[15] LTM processors make predictions and get feedback from CtmR’s inner and outer worlds. Based on this feedback, learning algorithms internal to each processor improve that processor’s behavior. These learning algorithms include each processor’s built-in Sleeping Experts Learning algorithm that helps the processor adjust weights it gives its gist. See (Blum A. , 1995) and (Blum, Hopcroft, & Kannan, 2015). [KM7] Thus, we can already see some basic predictive dynamics (prediction + testing + feedback + learning/correction) occurring within CtmR. [KM2] In time, some LTM processors become connected via bi-directional links.[16] Such links enable conscious communication, i.e. communication that goes through STM, to be replaced by more direct unconscious communication through links. Thus, when CtmR initially learns to ride a bike, most communication is done consciously until relevant processor links have formed. Then, for the most part, riding a bike is done unconsciously until an obstacle is encountered, forcing CtmR to pay conscious attention again. [KM6] CtmR’s competition, broadcast, attention and immediate direct communication via links is reminiscent of a process that Dehaene and Jean-Pierre Changeux call ignition (Dehaene & Changeux, 2005). But for feelings of consciousness, attention is not all you need. More is required. This paper is available on arxiv under CC BY 4.0 DEED license. [15] Thus, there is a (h+1) delay between the time the chunks are entered into the competition by the “unconscious” LTM processors and when CtmR pays conscious attention to the wining chunk. This is somewhat analogous to behavioral and brain studies going back to (Libet, 1985). [16] The CtmR has no links at birth. In our earlier “What’s her name?” scenario, when processor p sees that processors p’’ and p’’’ have useful information, p may form a bi-directional link with p’’ and perhaps also with p’’’. Authors: (1) Lenore Blum (lblum@cs.cmu.edu); (2) Manuel Blum (mblum@cs.cmu.edu). Authors: Authors: (1) Lenore Blum (lblum@cs.cmu.edu); (2) Manuel Blum (mblum@cs.cmu.edu). Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Brief Overview of CtmR, a Robot with a CTM Brain 2.1 Formal Definition of CtmR 2.1 Formal Definition of CtmR 2.2 Conscious Attention in CtmR 2.2 Conscious Attention in CtmR 2.3 Conscious Awareness and the Feeling of Consciousness in CtmR 2.3 Conscious Awareness and the Feeling of Consciousness in CtmR 2.4 CtmR as a Framework for Artificial General Intelligence (AGI) 2.4 CtmR as a Framework for Artificial General Intelligence (AGI) 3 Alignment of CtmR with Other Theories of Consciousness 3 Alignment of CtmR with Other Theories of Consciousness 4 Addressing Kevin Mitchell’s questions from the perspective of CtmR 4 Addressing Kevin Mitchell’s questions from the perspective of CtmR 5 Summary and Conclusions 5 Summary and Conclusions 6 Acknowledgements 6 Acknowledgements 7 Appendix 7.1 A Brief History of the Theoretical Computer Science Approach to Computation 7.1 A Brief History of the Theoretical Computer Science Approach to Computation 7.2 The Probabilistic Competition for Conscious Attention and the Influence of Disposition on it 7.2 The Probabilistic Competition for Conscious Attention and the Influence of Disposition on it References References 2.2 Conscious Attention in CtmR Formal definition 1. Conscious attention in CtmR is the reception by all LTM processors of the broadcast of CtmR’s current conscious content. [KM2] [KM5] Formal definition 1. In other words, CtmR pays conscious attention at time t+h+1 to the winner of the competition that commenced at time t.[15] LTM processors make predictions and get feedback from CtmR’s inner and outer worlds. Based on this feedback, learning algorithms internal to each processor improve that processor’s behavior. These learning algorithms include each processor’s built-in Sleeping Experts Learning algorithm that helps the processor adjust weights it gives its gist. See (Blum A. , 1995) and (Blum, Hopcroft, & Kannan, 2015). [KM7] Thus, we can already see some basic predictive dynamics (prediction + testing + feedback + learning/correction) occurring within CtmR. [KM2] In time, some LTM processors become connected via bi-directional links .[16] Such links enable conscious communication, i.e. communication that goes through STM, to be replaced by more direct unconscious communication through links. Thus, when CtmR initially learns to ride a bike, most communication is done consciously until relevant processor links have formed. Then, for the most part, riding a bike is done unconsciously until an obstacle is encountered, forcing CtmR to pay conscious attention again. [KM6] links CtmR’s competition, broadcast, attention and immediate direct communication via links is reminiscent of a process that Dehaene and Jean-Pierre Changeux call ignition (Dehaene & Changeux, 2005). But for feelings of consciousness, attention is not all you need. More is required. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv [15] Thus, there is a (h+1) delay between the time the chunks are entered into the competition by the “unconscious” LTM processors and when CtmR pays conscious attention to the wining chunk. This is somewhat analogous to behavioral and brain studies going back to (Libet, 1985). [16] The CtmR has no links at birth. In our earlier “What’s her name?” scenario, when processor p sees that processors p’’ and p’’’ have useful information, p may form a bi-directional link with p’’ and perhaps also with p’’’.

AT&T

att.com

10K+ employees

Since 1875

COMPANY RANKING

AT&T

EVERGREEN INDEX #3189

Representing this company? Click here to claim and customize this page!

AT&T WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

AT&T

att.com

10K+ employees

Since 1875

COMPANY RANKING

AT&T

EVERGREEN INDEX #3189

RELATED COMPANIESRANK

STORIES ABOUT AT&T

Representing this company? Click here to claim and customize this page!

AT&T WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!