Adi Shamir
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Adi Shamir has a total of 47 co-authors. Here are the top co-authors by number of shared publications:
- Orr Dunkelman
Faculty IV -- Electrical Engineering and Computer Science, Technische Universität Berlin, Berlin, Germany
12 shared publications - Nathan Keller
Department of Mathematics, Bar Ilan University, Ramat Gan, Israel
7 shared publications - Anna Hambitzer
Technology Innovation Institute
5 shared publications - Ben Nassi
Tel Aviv University, School of Electrical & Computer Engineering, Tel Aviv, Israel
4 shared publications - Eyal Ronen
Tel Aviv University, Tel Aviv, Israel
4 shared publications - Nathan Keller
Faculty of Mathematics and Computer Science, Weizmann Institute of Science, Rehovot, Israel
4 shared publications
Latest Publications
- 1 October 2026
- Published by Springer Nature in Journal of Cryptology
- Vol. 39 (4) , 32
Polynomial-Time Cryptanalytic Extraction of Deep Neural Networks in the Hard-Label Setting
- 1 October 2026
- Published by Springer Nature in Journal of Cryptology
- Vol. 39 (4) , 32
Polynomial-Time Cryptanalytic Extraction of Deep Neural Networks in the Hard-Label Setting
- Open Access
- 28 November 2025
- Published by Springer Nature in Discrete & Computational Geometry
- Vol. 75 (3) , 839-870
Error Resilient Space Partitioning
A major research area in discrete geometry is to consider the best way to partition the d -dimensional Euclidean space $$\mathbb {R}^d$$ under various quality criteria. In this paper we introduce a new type of space partitioning that is motivated by the problem of rounding noisy measurements from the continuous space $$\mathbb {R}^d$$ to a discrete subset of representative values. Specifically, we study partitions of $$\mathbb {R}^d$$ into bounded-size tiles colored by one of k colors, such that tiles of the same color have a distance of at least t from each other. Such tilings allow for error-resilient rounding, as two points of the same color and distance less than t from each other are guaranteed to belong to the same tile, and thus, to be rounded to the same point. The main problem we study in this paper is characterizing the achievable tradeoffs between the number of colors k and the distance t , for various dimensions d . On the qualitative side, we show that in $$\mathbb {R}^d$$ , using $$k=d+1$$ colors is both sufficient and necessary to achieve $$t>0$$ . On the quantitative side, we achieve numerous upper and lower bounds on t as a function of k . In particular, for $$d=3,4,8,24$$ , we obtain sharp asymptotic bounds on t , as $$k \rightarrow \infty $$ . We obtain our results with a variety of techniques including isoperimetric inequalities, the Brunn-Minkowski theorem, sphere packing bounds, Bapat’s connector-free lemma, and Čech cohomology.- 28 November 2025
- Published by Springer Nature in Discrete & Computational Geometry
- Vol. 75 (3) , 839-870
Open AccessError Resilient Space Partitioning
A major research area in discrete geometry is to consider the best way to partition the d -dimensional Euclidean space $$\mathbb {R}^d$$ under various quality criteria. In this paper we introduce a new type of space partitioning that is motivated by the problem of rounding noisy measurements from the continuous space $$\mathbb {R}^d$$ to a discrete subset of representative values. Specifically, we study partitions of $$\mathbb {R}^d$$ into bounded-size tiles colored by one of k colors, such that tiles of the same color have a distance of at least t from each other. Such tilings allow for error-resilient rounding, as two points of the same color and distance less than t from each other are guaranteed to belong to the same tile, and thus, to be rounded to the same point. The main problem we study in this paper is characterizing the achievable tradeoffs between the number of colors k and the distance t , for various dimensions d . On the qualitative side, we show that in $$\mathbb {R}^d$$ , using $$k=d+1$$ colors is both sufficient and necessary to achieve $$t>0$$ . On the quantitative side, we achieve numerous upper and lower bounds on t as a function of k . In particular, for $$d=3,4,8,24$$ , we obtain sharp asymptotic bounds on t , as $$k \rightarrow \infty $$ . We obtain our results with a variety of techniques including isoperimetric inequalities, the Brunn-Minkowski theorem, sphere packing bounds, Bapat’s connector-free lemma, and Čech cohomology. - 10 May 2025
- Published by Springer Nature
- p. 2678-2679
TWIRL
- 27 April 2025
- Published by Springer Nature
- p. 364-396
Polynomial Time Cryptanalytic Extraction of Deep Neural Networks in the Hard-Label Setting
- 27 April 2025
- Published by Springer Nature
- p. 364-396
Polynomial Time Cryptanalytic Extraction of Deep Neural Networks in the Hard-Label Setting
- Open Access
Polynomial Time Cryptanalytic Extraction of Deep Neural Networks in the Hard-Label Setting
Deep neural networks (DNNs) are valuable assets, yet their public accessibility raises security concerns about parameter extraction by malicious actors. Recent work by Carlini et al. (crypto'20) and Canales-Martínez et al. (eurocrypt'24) has drawn parallels between this issue and block cipher key extraction via chosen plaintext attacks. Leveraging differential cryptanalysis, they demonstrated that all the weights and biases of black-box ReLU-based DNNs could be inferred using a polynomial number of queries and computational time. However, their attacks relied on the availability of the exact numeric value of output logits, which allowed the calculation of their derivatives. To overcome this limitation, Chen et al. (asiacrypt'24) tackled the more realistic hard-label scenario, where only the final classification label (e.g., "dog" or "car") is accessible to the attacker. They proposed an extraction method requiring a polynomial number of queries but an exponential execution time. In addition, their approach was applicable only to a restricted set of architectures, could deal only with binary classifiers, and was demonstrated only on tiny neural networks with up to four neurons split among up to two hidden layers. This paper introduces new techniques that, for the first time, achieve cryptanalytic extraction of DNN parameters in the most challenging hard-label setting, using both a polynomial number of queries and polynomial time. We validate our approach by extracting nearly one million parameters from a DNN trained on the CIFAR-10 dataset, comprising 832 neurons in four hidden layers. Our results reveal the surprising fact that all the weights of a ReLU-based DNN can be efficiently determined by analyzing only the geometric shape of its decision boundaries.
Open AccessPolynomial Time Cryptanalytic Extraction of Deep Neural Networks in the Hard-Label Setting
Deep neural networks (DNNs) are valuable assets, yet their public accessibility raises security concerns about parameter extraction by malicious actors. Recent work by Carlini et al. (crypto'20) and Canales-Martínez et al. (eurocrypt'24) has drawn parallels between this issue and block cipher key extraction via chosen plaintext attacks. Leveraging differential cryptanalysis, they demonstrated that all the weights and biases of black-box ReLU-based DNNs could be inferred using a polynomial number of queries and computational time. However, their attacks relied on the availability of the exact numeric value of output logits, which allowed the calculation of their derivatives. To overcome this limitation, Chen et al. (asiacrypt'24) tackled the more realistic hard-label scenario, where only the final classification label (e.g., "dog" or "car") is accessible to the attacker. They proposed an extraction method requiring a polynomial number of queries but an exponential execution time. In addition, their approach was applicable only to a restricted set of architectures, could deal only with binary classifiers, and was demonstrated only on tiny neural networks with up to four neurons split among up to two hidden layers. This paper introduces new techniques that, for the first time, achieve cryptanalytic extraction of DNN parameters in the most challenging hard-label setting, using both a polynomial number of queries and polynomial time. We validate our approach by extracting nearly one million parameters from a DNN trained on the CIFAR-10 dataset, comprising 832 neurons in four hidden layers. Our results reveal the surprising fact that all the weights of a ReLU-based DNN can be efficiently determined by analyzing only the geometric shape of its decision boundaries.
- Open Access
- 15 July 2024
- Published by Springer Nature in Journal of Cryptology
- Vol. 37 (3) , 1-42
The Retracing Boomerang Attack, with Application to Reduced-Round AES
Boomerang attacks are extensions of differential attacks that make it possible to combine two unrelated differential properties of the first and second part of a cryptosystem with probabilities p and q into a new differential-like property of the whole cryptosystem with probability $$p^2q^2$$ (since each one of the properties has to be satisfied twice). In this paper, we describe a new version of boomerang attacks which uses the counterintuitive idea of throwing out most of the data in order to force equalities between certain values on the ciphertext side. In certain cases, this creates a correlation between the four probabilistic events, which increases the probability of the combined property to $$p^2q$$ and increases the signal-to-noise ratio of the resultant distinguisher. We call this variant a retracing boomerang attack since we make sure that the boomerang we throw follows the same path on its forward and backward directions. To demonstrate the power of the new technique, we apply it to the case of 5-round AES. This version of AES was repeatedly attacked by a large variety of techniques, but for twenty years its complexity had remained stuck at $$2^{32}$$ . At Crypto’18, it was finally reduced to $$2^{24}$$ (for full key recovery), and with our new technique, we can further reduce the complexity of full key recovery to the surprisingly low value of $$2^{16.5}$$ (i.e., only 90, 000 encryption/decryption operations are required for a full key recovery). In addition to improving previous attacks, our new technique unveils a hidden relationship between boomerang attacks and two other cryptanalytic techniques, the yoyo game and the recently introduced mixture differentials.- 15 July 2024
- Published by Springer Nature in Journal of Cryptology
- Vol. 37 (3) , 1-42
Open AccessThe Retracing Boomerang Attack, with Application to Reduced-Round AES
Boomerang attacks are extensions of differential attacks that make it possible to combine two unrelated differential properties of the first and second part of a cryptosystem with probabilities p and q into a new differential-like property of the whole cryptosystem with probability $$p^2q^2$$ (since each one of the properties has to be satisfied twice). In this paper, we describe a new version of boomerang attacks which uses the counterintuitive idea of throwing out most of the data in order to force equalities between certain values on the ciphertext side. In certain cases, this creates a correlation between the four probabilistic events, which increases the probability of the combined property to $$p^2q$$ and increases the signal-to-noise ratio of the resultant distinguisher. We call this variant a retracing boomerang attack since we make sure that the boomerang we throw follows the same path on its forward and backward directions. To demonstrate the power of the new technique, we apply it to the case of 5-round AES. This version of AES was repeatedly attacked by a large variety of techniques, but for twenty years its complexity had remained stuck at $$2^{32}$$ . At Crypto’18, it was finally reduced to $$2^{24}$$ (for full key recovery), and with our new technique, we can further reduce the complexity of full key recovery to the surprisingly low value of $$2^{16.5}$$ (i.e., only 90, 000 encryption/decryption operations are required for a full key recovery). In addition to improving previous attacks, our new technique unveils a hidden relationship between boomerang attacks and two other cryptanalytic techniques, the yoyo game and the recently introduced mixture differentials. - 1 May 2024
- Published by Springer Nature
- p. 3-33
Polynomial Time Cryptanalytic Extraction of Neural Network Models
- 1 May 2024
- Published by Springer Nature
- p. 3-33
Polynomial Time Cryptanalytic Extraction of Neural Network Models
- Open Access
Polynomial Time Cryptanalytic Extraction of Neural Network Models
Billions of dollars and countless GPU hours are currently spent on training Deep Neural Networks (DNNs) for a variety of tasks. Thus, it is essential to determine the difficulty of extracting all the parameters of such neural networks when given access to their black-box implementations. Many versions of this problem have been studied over the last 30 years, and the best current attack on ReLU-based deep neural networks was presented at Crypto 2020 by Carlini, Jagielski, and Mironov. It resembles a differential chosen plaintext attack on a cryptosystem, which has a secret key embedded in its black-box implementation and requires a polynomial number of queries but an exponential amount of time (as a function of the number of neurons). In this paper, we improve this attack by developing several new techniques that enable us to extract with arbitrarily high precision all the real-valued parameters of a ReLU-based DNN using a polynomial number of queries and a polynomial amount of time. We demonstrate its practical efficiency by applying it to a full-sized neural network for classifying the CIFAR10 dataset, which has 3072 inputs, 8 hidden layers with 256 neurons each, and over million neuronal parameters. An attack following the approach by Carlini et al. requires an exhaustive search over 2 to the power 256 possibilities. Our attack replaces this with our new techniques, which require only 30 minutes on a 256-core computer.
Open AccessPolynomial Time Cryptanalytic Extraction of Neural Network Models
Billions of dollars and countless GPU hours are currently spent on training Deep Neural Networks (DNNs) for a variety of tasks. Thus, it is essential to determine the difficulty of extracting all the parameters of such neural networks when given access to their black-box implementations. Many versions of this problem have been studied over the last 30 years, and the best current attack on ReLU-based deep neural networks was presented at Crypto 2020 by Carlini, Jagielski, and Mironov. It resembles a differential chosen plaintext attack on a cryptosystem, which has a secret key embedded in its black-box implementation and requires a polynomial number of queries but an exponential amount of time (as a function of the number of neurons). In this paper, we improve this attack by developing several new techniques that enable us to extract with arbitrarily high precision all the real-valued parameters of a ReLU-based DNN using a polynomial number of queries and a polynomial amount of time. We demonstrate its practical efficiency by applying it to a full-sized neural network for classifying the CIFAR10 dataset, which has 3072 inputs, 8 hidden layers with 256 neurons each, and over million neuronal parameters. An attack following the approach by Carlini et al. requires an exhaustive search over 2 to the power 256 possibilities. Our attack replaces this with our new techniques, which require only 30 minutes on a 256-core computer.
- 10 October 2023
- Published by Springer Nature in Designs, Codes and Cryptography
- Vol. 92 (1) , 159-177
Quantum time/memory/data tradeoff attacks
- 10 October 2023
- Published by Springer Nature in Designs, Codes and Cryptography
- Vol. 92 (1) , 159-177
Quantum time/memory/data tradeoff attacks
- 15 April 2023
- Published by Springer Nature
- p. 98-127
Efficient Detection of High Probability Statistical Properties of Cryptosystems via Surrogate Differentiation
- 15 April 2023
- Published by Springer Nature
- p. 98-127
Efficient Detection of High Probability Statistical Properties of Cryptosystems via Surrogate Differentiation