Equalized odds, also referred to as conditional procedure accuracy equality and disparate mistreatment, is a measure of fairness in machine learning. A classifier satisfies this definition if the subjects in the protected and unprotected groups have equal true positive rate and equal false positive rate, satisfying the formula: P ( R = + | Y = y , A = a ) = P ( R = + | Y = y , A = b ) y ∈ { + , − } ∀ a , b ∈ A {\displaystyle P(R=+|Y=y,A=a)=P(R=+|Y=y,A=b)\quad y\in \{+,-\}\quad \forall a,b\in A} For example, A {\displaystyle A} could be gender, race, or any other characteristics that we want to be free of bias, while Y {\displaystyle Y} would be whether the person is qualified for the degree, and the output R {\displaystyle R} would be the school's decision whether to offer the person to study for the degree. In this context, higher university enrollment rates of African Americans compared to whites with similar test scores might be necessary to fulfill the condition of equalized odds, if the "base rate" of Y {\displaystyle Y} differs between the groups. The concept was originally defined for binary-valued Y {\displaystyle Y} . In 2017, Woodworth et al. generalized the concept further for multiple classes.
System integrity
In telecommunications, the term system integrity has the following meanings: That condition of a system wherein its mandated operational and technical parameters are within the prescribed limits. The quality of an AIS when it performs its intended function in an unimpaired manner, free from deliberate or inadvertent unauthorized manipulation of the system. The state that exists when there is complete assurance that under all conditions an IT system is based on the logical correctness and reliability of the operating system, the logical completeness of the hardware and software that implement the protection mechanisms, and data integrity.
Forking lemma
The forking lemma is any of a number of related lemmas in cryptography research. The lemma states that if an adversary (typically a probabilistic Turing machine), on inputs drawn from some distribution, produces an output that has some property with non-negligible probability, then with non-negligible probability, if the adversary is re-run on new inputs but with the same random tape, its second output will also have the property. This concept was first used by David Pointcheval and Jacques Stern in "Security proofs for signature schemes," published in the proceedings of Eurocrypt 1996. In their paper, the forking lemma is specified in terms of an adversary that attacks a digital signature scheme instantiated in the random oracle model. They show that if an adversary can forge a signature with non-negligible probability, then there is a non-negligible probability that the same adversary with the same random tape can create a second forgery in an attack with a different random oracle. The forking lemma was later generalized by Mihir Bellare and Gregory Neven. The forking lemma has been used and further generalized to prove the security of a variety of digital signature schemes and other random-oracle based cryptographic constructions. == Statement of the lemma == The generalized version of the lemma is stated as follows. Let A be a probabilistic algorithm, with inputs (x, h1, ..., hq; r) that outputs a pair (J, y), where r refers to the random tape of A (that is, the random choices A will make). Suppose further that IG is a probability distribution from which x is drawn, and that H is a set of size h from which each of the hi values are drawn according to the uniform distribution. Let acc be the probability that on inputs distributed as described, the J output by A is greater than or equal to 1. We can then define a "forking algorithm" FA that proceeds as follows, on input x: Pick a random tape r for A. Pick h1, ..., hq uniformly from H. Run A on input (x, h1, ..., hq; r) to produce (J, y). If J = 0, then return (0, 0, 0). Pick h'J, ..., h'q uniformly from H. Run A on input (x, h1, ..., hJ−1, h'J, ..., h'q; r) to produce (J', y'). If J' = J and hJ ≠ h'J then return (1, y, y'), otherwise, return (0, 0, 0). Let frk be the probability that FA outputs a triple starting with 1, given an input x chosen randomly from IG. Then frk ≥ acc ⋅ ( acc q − 1 h ) . {\displaystyle {\text{frk}}\geq {\text{acc}}\cdot \left({\frac {\text{acc}}{q}}-{\frac {1}{h}}\right).} === Intuition === The idea here is to think of A as running two times in related executions, where the process "forks" at a certain point, when some but not all of the input has been examined. In the alternate version, the remaining inputs are re-generated but are generated in the normal way. The point at which the process forks may be something we only want to decide later, possibly based on the behavior of A the first time around: this is why the lemma statement chooses the branching point (J) based on the output of A. The requirement that hJ ≠ h'J is a technical one required by many uses of the lemma. (Note that since both hJ and h'J are chosen randomly from H, then if h is large, as is usually the case, the probability of the two values not being distinct is extremely small.) === Example === For example, let A be an algorithm for breaking a digital signature scheme in the random oracle model. Then x would be the public parameters (including the public key) A is attacking, and hi would be the output of the random oracle on its ith distinct input. The forking lemma is of use when it would be possible, given two different random signatures of the same message, to solve some underlying hard problem. An adversary that forges once, however, gives rise to one that forges twice on the same message with non-negligible probability through the forking lemma. When A attempts to forge on a message m, we consider the output of A to be (J, y) where y is the forgery, and J is such that m was the Jth unique query to the random oracle (it may be assumed that A will query m at some point, if A is to be successful with non-negligible probability). (If A outputs an incorrect forgery, we consider the output to be (0, y).) By the forking lemma, the probability (frk) of obtaining two good forgeries y and y' on the same message but with different random oracle outputs (that is, with hJ ≠ h'J) is non-negligible when acc is also non-negligible. This allows us to prove that if the underlying hard problem is indeed hard, then no adversary can forge signatures. This is the essence of the proof given by Pointcheval and Stern for a modified ElGamal signature scheme against an adaptive adversary. == Known issues with application of forking lemma == The reduction provided by the forking lemma is not tight. Pointcheval and Stern proposed security arguments for Digital Signatures and Blind Signature using Forking Lemma. Claus P. Schnorr provided an attack on blind Schnorr signatures schemes, with more than p o l y l o g ( n ) {\displaystyle polylog(n)} concurrent executions (the case studied and proven secure by Pointcheval and Stern). A polynomial-time attack, for Ω ( n ) {\displaystyle \Omega (n)} concurrent executions, was shown in 2020 by Benhamouda, Lepoint, Raykova, and Orrù. Schnorr also suggested enhancements for securing blind signatures schemes based on discrete logarithm problem.
Data exhaust
Data exhaust (also exhaust data) is the trail of data generated as a by-product of users' online activity, behaviour, and transactions, rather than data they deliberately create or submit. It forms part of a broader category of unconventional data that also includes geospatial, network, and time-series data, and may be useful for predictive analytics. Data exhaust can take the form of cookies, temporary files, log files, clickstream records and stored preferences. Actions such as visiting a web page, following a link, or dwelling on an element may all generate exhaust data that is recorded without the user's active awareness. Unlike primary content — which the user intentionally creates — exhaust data is a passive side effect of interaction. A bank, for example, might treat the amounts and parties involved in a transaction as primary data, while secondary data could include whether the transaction was carried out at a cash machine rather than a branch. == Uses == Data exhaust collected by companies is often information that is not immediately useful in isolation, but can be aggregated and analysed to improve products, personalise content, identify trends, and support quality control. Companies may also store exhaust data for future analysis or sell it to third parties. Shoshana Zuboff has described this practice as a core mechanism of what she terms surveillance capitalism, in which behavioural data generated by users is converted into predictive products. Kosciejew notes that large quantities of often raw data are collected in this way, much of which is never analysed. == Medical exhaust data == Many medical devices — including pacemakers, dialysis machines and surgical cameras — generate exhaust data as a by-product of their operation. The majority of this data is never captured or analysed, and is typically discarded once a procedure ends or a device completes its routine monitoring cycle. The potential use of data generated by implanted devices such as pacemakers raises additional legal and ethical questions around ownership and consent. Using electronic health records for research also creates challenges because of the volume of data involved, creating a need for automated algorithms to process it. == Privacy and regulation == The collection and distribution of data exhaust is not in itself illegal in most jurisdictions, but its use raises questions of privacy and informed consent. Steps commonly taken to address these concerns include data anonymisation, offering users an opt-out from the sale of their data, and publishing explicit privacy policies that disclose what data is collected and how it is used.
Computer network
In computer science, computer engineering, and telecommunications, a network is a group of communicating computers and peripherals known as hosts, which communicate data to other hosts via communication protocols, as facilitated by networking hardware. Within a computer network, hosts are identified by network addresses, which allow networking hardware to locate and identify hosts. Hosts may also have hostnames, memorable labels for the host nodes, which can be mapped to a network address using a hosts file or a name server such as Domain Name Service. The physical medium that supports information exchange includes wired media like copper cables, optical fibers, and wireless radio-frequency media. The arrangement of hosts and hardware within a network architecture is known as the network topology. The first computer network was created in 1940 when George Stibitz connected a terminal at Dartmouth to his Complex Number Calculator at Bell Labs in New York. Today, almost all computers are connected to a computer network, such as the global Internet or embedded networks such as those found in many modern electronic devices. Many applications have only limited functionality unless they are connected to a network. Networks support applications and services, such as access to the World Wide Web, digital video and audio, application and storage servers, printers, and email and instant messaging applications. == History == === Early origins (1940 – 1960s) === In 1940, George Stibitz of Bell Labs connected a teletype at Dartmouth to a Bell Labs computer running his Complex Number Calculator to demonstrate the use of computers at long distance. This was the first real-time, remote use of a computing machine. In the late 1950s, a network of computers was built for the U.S. military Semi-Automatic Ground Environment (SAGE) radar system using the Bell 101 modem. It was the first commercial modem for computers, released by AT&T Corporation in 1958. The modem allowed digital data to be transmitted over regular unconditioned telephone lines at a speed of 110 bits per second (bit/s). In 1959, Christopher Strachey filed a patent application for time-sharing in the United Kingdom and John McCarthy initiated the first project to implement time-sharing of user programs at MIT. Strachey passed the concept on to J. C. R. Licklider at the inaugural UNESCO Information Processing Conference in Paris that year. McCarthy was instrumental in the creation of three of the earliest time-sharing systems (the Compatible Time-Sharing System in 1961, the BBN Time-Sharing System in 1962, and the Dartmouth Time-Sharing System in 1963). In 1959, Anatoly Kitov proposed to the Central Committee of the Communist Party of the Soviet Union a detailed plan for the re-organization of the control of the Soviet armed forces and of the Soviet economy on the basis of a network of computing centers. Kitov's proposal was rejected, as later was the 1962 OGAS economy management network project. During the 1960s, Paul Baran and Donald Davies independently invented the concept of packet switching for data communication between computers over a network. Baran's work addressed adaptive routing of message blocks across a distributed network, but did not include routers with software switches, nor the idea that users, rather than the network itself, would provide the reliability. Davies' hierarchical network design included high-speed routers, communication protocols and the essence of the end-to-end principle. The NPL network, a local area network at the National Physical Laboratory (United Kingdom), pioneered the implementation of the concept in 1968-69 using 768 kbit/s links. Both Baran's and Davies' inventions were seminal contributions that influenced the development of computer networks. === ARPANET (1969 – 1974) === In 1962 and 1963, J. C. R. Licklider sent a series of memos to office colleagues discussing the concept of the "Intergalactic Computer Network", a computer network intended to allow general communications among computer users. This ultimately became the basis for the ARPANET, which began in 1969. That year, the first four nodes of the ARPANET were connected using 50 kbit/s circuits between the University of California at Los Angeles, the Stanford Research Institute, the University of California, Santa Barbara, and the University of Utah. Designed principally by Bob Kahn, the network's routing, flow control, software design and network control were developed by the IMP team working for Bolt Beranek & Newman. In the early 1970s, Leonard Kleinrock carried out mathematical work to model the performance of packet-switched networks, which underpinned the development of the ARPANET. His theoretical work on hierarchical routing in the late 1970s with student Farouk Kamoun remains critical to the operation of the Internet today. In 1973, Peter Kirstein put internetworking into practice at University College London (UCL), connecting the ARPANET to British academic networks, the first international heterogeneous computer network. That same year, Robert Metcalfe wrote a formal memo at Xerox PARC describing Ethernet, a local area networking system he created with David Boggs. It was inspired by the packet radio ALOHAnet, started by Norman Abramson and Franklin Kuo at the University of Hawaii in the late 1960s. Metcalfe and Boggs, with John Shoch and Edward Taft, also developed the PARC Universal Packet for internetworking. That year, the French CYCLADES network, directed by Louis Pouzin was the first to make the hosts responsible for the reliable delivery of data, rather than this being a centralized service of the network itself. === The internet (1974 – present) === In 1974, Vint Cerf and Bob Kahn published their seminal 1974 paper on internetworking, A Protocol for Packet Network Intercommunication. Later that year, Cerf, Yogen Dalal, and Carl Sunshine wrote the first Transmission Control Protocol (TCP) specification, RFC 675, coining the term Internet as a shorthand for internetworking. In July 1976, Metcalfe and Boggs published their paper "Ethernet: Distributed Packet Switching for Local Computer Networks" and in December 1977, together with Butler Lampson and Charles P. Thacker, they received U.S. patent 4063220A for their invention. In 1976, John Murphy of Datapoint Corporation created ARCNET, a token-passing network first used to share storage devices. In 1979, Robert Metcalfe pursued making Ethernet an open standard. In 1980, Ethernet was upgraded from the original 2.94 Mbit/s protocol to the 10 Mbit/s protocol, which was developed by Ron Crane, Bob Garner, Roy Ogus, Hal Murray, Dave Redell and Yogen Dalal. In 1986, the National Science Foundation (NSF) launched the National Science Foundation Network (NSFNET) as a general-purpose research network connecting various NSF-funded sites to each other and to regional research and education networks. In 1995, the transmission speed capacity for Ethernet increased from 10 Mbit/s to 100 Mbit/s. By 1998, Ethernet supported transmission speeds of 1 Gbit/s. Subsequently, higher speeds of up to 800 Gbit/s were added (as of 2025). The scaling of Ethernet has been a contributing factor to its continued use. In the 1980s and 1990s, as embedded systems were becoming increasingly important in factories, cars, and airplanes, network protocols were developed to allow the embedded computers to communicate. In the late 1990s and 2000s, ubiquitous computing and an Internet of Things became popular. === Commercial usage === In 1960, the commercial airline reservation system semi-automatic business research environment (SABRE) went online with two connected mainframes. In 1965, Western Electric introduced the first widely used telephone switch that implemented computer control in the switching fabric. In 1972, commercial services were first deployed on experimental public data networks in Europe. Public data networks in Europe, North America and Japan began using X.25 in the late 1970s and interconnected with X.75. This underlying infrastructure was used for expanding TCP/IP networks in the 1980s. In 1977, the first long-distance fiber network was deployed by GTE in Long Beach, California. == Hardware == === Network links === The transmission media used to link devices to form a computer network include electrical cable, optical fiber, and free space. In the OSI model, the software to handle the media is defined at layers 1 and 2 — the physical layer and the data link layer. Common examples of networking technologies include: Ethernet is a widely adopted family of networking technologies that use copper and fiber media in local area networks (LAN). The media and protocol standards that enable communication between networked devices over Ethernet are defined by IEEE 802.3. Wireless LAN standards, which use radio waves. Some standards use infrared signals as a transmission medium. Power line communication uses a building's power cabling to transmit
Artificial intelligence content detection
Artificial intelligence detection software aims to determine whether some content (text, image, video, or audio) was generated using artificial intelligence (AI). This software is often unreliable. == Accuracy issues == Many AI detection tools have been shown to be unreliable in detecting AI-generated text. In a 2023 study conducted by Weber-Wulff et al., researchers evaluated 14 detection tools including Turnitin and GPTZero and found that "all scored below 80% of accuracy and only 5 over 70%." They also found that these tools tend to have a bias for classifying texts more as human than as AI, and that accuracy of these tools worsens upon paraphrasing. === False positives === In AI content detection, a false positive is when human-written work is incorrectly flagged as AI-written. Many AI detection platforms claim to have a minimal level of false positives, with Turnitin claiming a less than 1% false positive rate. However, later research by The Washington Post produced much higher rates of 50%, though they used a smaller sample size. False positives in an academic setting frequently lead to accusations of academic misconduct, which can have serious consequences for a student's academic record. Additionally, studies have shown evidence that many AI detection models are prone to give false positives to work written by people whose first language is not English, and also to neurodivergent people. In June 2023, Janelle Shane wrote that portions of her book You Look Like a Thing and I Love You were flagged as AI-generated. === False negatives === A false negative is a failure to identify documents with AI-written text. False negatives often happen as a result of a detection software's sensitivity level or because evasive techniques were used when generating the work to make it sound more human. False negatives are less of a concern academically, since they aren't likely to lead to accusations and ramifications. Notably, Turnitin stated they have a 15% false negative rate. == Text detection == For text, this is usually done to prevent alleged plagiarism, often by detecting repetition of words as telltale signs that a text was AI-generated (including hallucinations). Detection systems may also rely on stylistic and structural regularities associated with LLM output, such as unusually consistent grammar, formulaic transitions, repeated discourse markers, and recurring rhetorical templates. Some tools are designed less to establish authorship provenance than to flag prose that resembles common LLM-generated style patterns. They are often used by teachers marking their students, usually on an ad hoc basis. Following the release of ChatGPT and similar AI text generative software, many educational establishments have issued policies against the use of AI by students. AI text detection software is also used by those assessing job applicants, as well as online search engines, hiring, online moderation and publishing. Current detectors may sometimes be unreliable and have incorrectly marked work by humans as originating from AI while failing to detect AI-generated work in other instances. MIT Technology Review said that the technology "struggled to pick up ChatGPT-generated text that had been slightly rearranged by humans and obfuscated by a paraphrasing tool". AI text detection software has also been shown to discriminate against non-native speakers of English. Two students from the University of California, Davis, were referred to the university's Office of Student Success and Judicial Affairs (OSSJA) after their professors scanned their essays with positive results; the first with an AI detector called GPTZero, and the second with an AI detector integration in Turnitin. However, following media coverage, and a thorough investigation, the students were cleared of any wrongdoing. In April 2023, Cambridge University and other members of the Russell Group of universities in the United Kingdom opted out of Turnitin's AI text detection tool, after expressing concerns it was unreliable. The University of Texas at Austin opted out of the system six months later. In May 2023, a professor at Texas A&M University–Commerce used ChatGPT to detect whether his students' content was written by it, which ChatGPT said was the case. As such, he threatened to fail the class despite ChatGPT not being able to detect AI-generated writing. No students were prevented from graduating because of the issue, and all but one student (who admitted to using the software) were exonerated from accusations of having used ChatGPT in their content. In July 2023, a paper titled "GPT detectors are biased against non-native English writers" was released, reporting that GPTs discriminate against non-native English authors. The paper compared seven GPT detectors against essays from both non-native English speakers and essays from United States students. The essays from non-native English speakers had an average false positive rate of 61.3%. An article by Thomas Germain, published on Gizmodo in June 2024, reported job losses among freelance writers and journalists due to AI text detection software mistakenly classifying their work as AI-generated. In September 2024, Common Sense Media reported that generative AI detectors had a 20% false positive rate for Black students, compared to 10% of Latino students and 7% of White students. To improve the reliability of AI text detection, researchers have explored digital watermarking techniques. A 2023 paper titled "A Watermark for Large Language Models" presents a method to embed imperceptible watermarks into text generated by large language models (LLMs). This watermarking approach allows content to be flagged as AI-generated with a high level of accuracy, even when text is slightly paraphrased or modified. The technique is designed to be subtle and hard to detect for casual readers, thereby preserving readability, while providing a detectable signal for those employing specialized tools. However, while promising, watermarking faces challenges in remaining robust under adversarial transformations and ensuring compatibility across different LLMs. == Anti text detection == There is software available designed to bypass AI text detection. In practice, evasion may not require specialized bypass tools. Paraphrasing, style editing, and removal of repeated discourse markers can substantially reduce the effectiveness of detectors that rely on recognizable surface patterns. A study published in August 2023 analyzed 20 abstracts from papers published in the Eye Journal, which were then paraphrased using GPT-4.0. The AI-paraphrased abstracts were examined for plagiarism using QueText and for AI-generated content using Originality.AI. The texts were then re-processed through an adversarial software called Undetectable.ai in order to reduce the AI-detection scores. The study found that the AI detection tool, Originality.AI, identified text generated by GPT-4 with a mean accuracy of 91.3%. However, after reprocessing by Undetectable.ai, the detection accuracy of Originality.ai dropped to a mean accuracy of 27.8%. Some experts also believe that techniques like digital watermarking are ineffective because they can be removed or added to trigger false positives. "A Watermark for Large Language Models" paper by Kirchenbauer et al. (2023) also addresses potential vulnerabilities of watermarking techniques. The authors outline a range of adversarial tactics, including text insertion, deletion, and substitution attacks, that could be used to bypass watermark detection. These attacks vary in complexity, from simple paraphrasing to more sophisticated approaches involving tokenization and homoglyph alterations. The study highlights the challenge of maintaining watermark robustness against attackers who may employ automated paraphrasing tools or even specific language model replacements to alter text spans iteratively while retaining semantic similarity. Experimental results show that although such attacks can degrade watermark strength, they also come at the cost of text quality and increased computational resources. == Image, video, and audio detection == Several purported AI image detection software exist, to detect AI-generated images (for example, those originating from Midjourney or DALL-E). They are not completely reliable. Industry analyses have also noted that AI-driven image recognition systems often struggle in real-world environments, where inconsistent lighting, noise and variable visual inputs reduce detection reliability, a challenge highlighted in modern agricultural quality-control research. Others claim to identify video and audio deepfakes, but this technology is also not fully reliable yet either. Despite debate around the efficacy of watermarking, Google DeepMind is actively developing a detection software called SynthID, which works by inserting a digital watermark that is invisible to the human eye into the pixels of an image.
White-box cryptography
In cryptography, the white-box model refers to an extreme attack scenario, in which an adversary has full unrestricted access to a cryptographic implementation, most commonly of a block cipher such as the Advanced Encryption Standard (AES). A variety of security goals may be posed (see the section below), the most fundamental being "unbreakability", requiring that any (bounded) attacker should not be able to extract the secret key hardcoded in the implementation, while at the same time the implementation must be fully functional. In contrast, the black-box model only provides an oracle access to the analyzed cryptographic primitive (in the form of encryption and/or decryption queries). There is also a model in-between, the so-called gray-box model, which corresponds to additional information leakage from the implementation, more commonly referred to as side-channel leakage. White-box cryptography is a practice and study of techniques for designing and attacking white-box implementations. It has many applications, including digital rights management (DRM), pay television, protection of cryptographic keys in the presence of malware, mobile payments and cryptocurrency wallets. Examples of DRM systems employing white-box implementations include CSS and Widevine. White-box cryptography is closely related to the more general notions of obfuscation, in particular, to Black-box obfuscation, proven to be impossible, and to Indistinguishability obfuscation, constructed recently under well-founded assumptions but so far being infeasible to implement in practice. As of January 2023, there are no publicly known unbroken white-box designs of standard symmetric encryption schemes. On the other hand, there exist many unbroken white-box implementations of dedicated block ciphers designed specifically to achieve incompressibility (see § Security goals). == Security goals == Depending on the application, different security goals may be required from a white-box implementation. Specifically, for symmetric-key algorithms the following are distinguished: Unbreakability is the most fundamental goal requiring that a bounded attacker should not be able to recover the secret key embedded in the white-box implementation. Without this requirement, all other security goals are unreachable since a successful attacker can simply use a reference implementation of the encryption scheme together with the extracted key. One-wayness requires that a white-box implementation of an encryption scheme can not be used by a bounded attacker to decrypt ciphertexts. This requirement essentially turns a symmetric encryption scheme into a public-key encryption scheme, where the white-box implementation plays the role of the public key associated to the embedded secret key. This idea was proposed already in the famous work of Diffie and Hellman in 1976 as a potential public-key encryption candidate. Code lifting security is an informal requirement on the context, in which the white-box program is being executed. It demands that an attacker can not extract a functional copy of the program. This goal is particularly relevant in the DRM setting. Code obfuscation techniques are often used to achieve this goal. A commonly used technique is to compose the white-box implementation with so-called external encodings. These are lightweight secret encodings that modify the function computed by the white-box part of an application. It is required that their effect is canceled in other parts of the application in an obscure way, using code obfuscation techniques. Alternatively, the canceling counterparts can be applied on a remote server. Incompressibility requires that an attacker can not significantly compress a given white-box implementation. This can be seen as a way to achieve code lifting security (see above), since exfiltrating a large program from a constrained device (for example, an embedded or a mobile device) can be time-consuming and may be easy to detect by a firewall. Examples of incompressible designs include SPACE cipher, SPNbox, WhiteKey and WhiteBlock. These ciphers use large lookup tables that can be pseudorandomly generated from a secret master key. Although this makes the recovery of the master key hard, the lookup tables themselves play the role of an equivalent secret key. Thus, unbreakability is achieved only partially. Traceability (Traitor tracing) requires that each distributed white-box implementation contains a digital watermark allowing identification of the guilty user in case the white-box program is being leaked and distributed publicly. == History == The white-box model with initial attempts of white-box DES and AES implementations were first proposed by Chow, Eisen, Johnson and van Oorshot in 2003. The designs were based on representing the cipher as a network of lookup tables and obfuscating the tables by composing them with small (4- or 8-bit) random encodings. Such protection satisfied a property that each single obfuscated table individually does not contain any information about the secret key. Therefore, a potential attacker has to combine several tables in their analysis. The first two schemes were broken in 2004 by Billet, Gilbert, and Ech-Chatbi using structural cryptanalysis. The attack was subsequently called "the BGE attack". The numerous consequent design attempts (2005-2022) were quickly broken by practical dedicated attacks. In 2016, Bos, Hubain, Michiels and Teuwen showed that an adaptation of standard side-channel power analysis attacks can be used to efficiently and fully automatically break most existing white-box designs. This result created a new research direction about generic attacks (correlation-based, algebraic, fault injection) and protections against them. == Competitions == Four editions of the WhibOx contest were held in 2017, 2019, 2021 and 2024 respectively. These competitions invited white-box designers both from academia and industry to submit their implementation in the form of (possibly obfuscated) C code. At the same time, everyone could attempt to attack these programs and recover the embedded secret key. Each of these competitions lasted for about 4-5 months. WhibOx 2017 / CHES 2017 Capture the Flag Challenge targeted the standard AES block cipher. Among 94 submitted implementations, all were broken during the competition, with the strongest one staying unbroken for 28 days. WhibOx 2019 / CHES 2019 Capture the Flag Challenge again targeted the AES block cipher. Among 27 submitted implementations, 3 programs stayed unbroken throughout the competition, but were broken after 51 days since the publication. WhibOx 2021 / CHES 2021 Capture the Flag Challenge changed the target to ECDSA, a digital signature scheme based on elliptic curves. Among 97 submitted implementations, all were broken within at most 2 days. WhibOx 2024 / CHES 2024 Capture the Flag Challenge again targeted ECDSA. Among 47 submitted implementations, all were broken during the competition, with the strongest one staying unbroken for almost 5 days.