The Hidden Parallelepiped Is Back Again: Power Analysis Attacks on Falcon
Falcon is a very efficient and compact lattice-based signature finalist of the NIST's Post-Quantum standardization campaign. This work assesses Falcon's side-channel resistance by analyzing two vulnerabilities, namely the pre-image computation and the trapdoor sampling. The first attack is an improvement of Karabulut and Aysu (DAC 2021). It overcomes several difficulties inherent to the structure of the stored key like the Fourier representation and directly recovers the key with a limited number of traces and a reduced complexity. The main part of this paper is dedicated to our second attack: we show that a simple power analysis during the signature execution could provide the exact value of the output of a subroutine called the base sampler. This intermediate value does not directly lead to the secret and we had to adapt the so-called hidden parallelepiped attack initially introduced by Nguyen and Regev in Eurocrypt 2006 and reused by Ducas and Nguyen in Asiacrypt 2012. We extensively quantify the resources for our attacks and experimentally demonstrate them with Falcon's reference implementation on the ELMO simulator (McCann, Oswald and Whitnall USENIX 2017) and on a ChipWhisperer Lite with STM32F3 target (ARM Cortex M4). These new attacks highlight the need for side-channel protection for one of the three finalists of NIST's standardization campaign by pointing out the vulnerable parts and quantifying the resources of the attacks.
Unifying Leakage Models on a Rényi Day 📺
In the last decade, several works have focused on finding the best way to model the leakage in order to obtain provably secure implementations. One of the most realistic models is the noisy leakage model, introduced in [PR13, DDF14] together with secure constructions. These works suffer from various limitations, in particular the use of ideal leak-free gates in [PR13] and an important loss (in the size of the field) in the reduction in [DDF14].In this work, we provide new strategies to prove the security of masked implementations and start by unifying the different noisiness metrics used in prior works by relating all of them to a standard notion in information theory: the pointwise mutual information. Based on this new interpretation, we define two new natural metrics and analyze the security of known compilers with respect to these metrics. In particular, we prove (1) a tighter bound for reducing the noisy leakage models to the probing model using our first new metric, (2) better bounds for amplification-based security proofs using the second metric.To support that the improvements we obtain are not only a consequence of the use of alternative metrics, we show that for concrete representation of leakage (e.g., “Hamming weight + Gaussian noise”), our approach significantly improves the parameters compared to prior works. Finally, using the Rényi divergence, we quantify concretely the advantage of an adversary in attacking a block cipher depending on the number of leakage acquisitions available to it.