Search

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-Mart\'inez 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.

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.

In SIDH and SIKE protocols, public keys are defined over quadratic extensions of prime fields. We present in this work a projective invariant property characterizing affine Montgomery curves defined over prime fields. We then force a secret 3-isogeny chain to repeatedly pass through a curve defined over a prime field in order to exploit the new property and inject zeros in the A-coefficient of an intermediate curve to successfully recover the isogeny chain one step at a time. Our results introduce a new kind of fault attacks applicable to SIDH and SIKE., Comment: Update: extended countermeasures

Let $q = 2^n$, and let $E / \mathbb{F}_{q^{\ell}}$ be a generalized Galbraith--Lin--Scott (GLS) binary curve, with $\ell \ge 2$ and $(\ell, n) = 1$.We show that the GLS endomorphism on $E / \mathbb{F}_{q^{\ell}}$ induces an efficient endomorphism on the Jacobian $J_H(\mathbb{F}_q)$ of the genus-$g$ hyperelliptic curve $H$ corresponding to the image of the GHS Weil-descent attack applied to $E/\mathbb{F}_{q^\ell}$, and that this endomorphism yields a factor-$n$ speedup when using standard index-calculus procedures for solving the Discrete Logarithm Problem (DLP) on $J_H(\mathbb{F}_q)$. Our analysis is backed up by the explicit computation of a discrete logarithm defined on a prime-order subgroup of a GLS elliptic curve over the field $\mathbb{F}_{2^{5\cdot 31}}$. A Magma implementation of our algorithm finds the aforementioned discrete logarithm in about $1,035$ CPU-days., Comment: Finite Fields and Their Applications, Elsevier, In press, 75

CSIDH is a recent quantum-resistant primitive based on the difficulty of finding isogeny paths between supersingular curves. Recently, two constant-time versions of CSIDH have been proposed: first by Meyer, Campos and Reith, and then by Onuki, Aikawa, Yamazaki and Takagi. While both offer protection against timing attacks and simple power consumption analysis, they are vulnerable to more powerful attacks such as fault injections. In this work, we identify and repair two oversights in these algorithms that compromised their constant-time character. By exploiting Edwards arithmetic and optimal addition chains, we produce the fastest constant-time version of CSIDH to date. We then consider the stronger attack scenario of fault injection, which is relevant for the security of CSIDH static keys in embedded hardware. We propose and evaluate a dummy-free CSIDH algorithm. While these CSIDH variants are slower, their performance is still within a small constant factor of less-protected variants. Finally, we discuss derandomized CSIDH algorithms., Comment: This work has been accepted in LATINCRYPT-2019

Let $f:\mathbb{C}\sp n\to\mathbb{C}\sp n$, $n\geq2$, be a biholomorphism and let $\Lambda\subseteq \mathbb{C}\sp n$ be a compact $f$-invariant set such that $f|\Lambda$ is partially hyperbolic. We give equivalent conditions to hyperbolicity on $\Lambda$. In the particular case of generalized H\'enon map with dominated splitting in the Julia set $J$, we characterize the hyperbolicity of $J$.

Dose volume histograms are a useful tool in state-of-the-art radiotherapy planning, and it is essential to be aware of their limitations. Dose distributions computed by treatment planning systems are affected by several sources of uncertainty such as algorithm limitations, measurement uncertainty in the data used to model the beam and residual differences between measured and computed dose, once the model is optimized. In order to take into account the effect of uncertainty, a probabilistic approach is proposed and a new kind of histogram, a dose-expected volume histogram, is introduced. The expected value of the volume in the region of interest receiving an absorbed dose equal or greater than a certain value is found using the probability distribution of the dose at each point. A rectangular probability distribution is assumed for this point dose, and a relationship is given for practical computations. This method is applied to a set of dose volume histograms for different regions of interest for 6 brain patients, 8 lung patients, 8 pelvis patients and 6 prostate patients planned for IMRT. These results show how dose computation uncertainty has effects on PTV coverage and, to a lesser extent, in dose to organs at risk. This method allows to quantify these effects., Comment: 17 pages, 10 figures, 5 tables. Submitted