DSCAD: Dilithium Side-Channel Attacks Dataset

Introduction：There is a lack of publicly available and standardized side-channel datasets to support the practical application of deep learning methods. This gap severely limits the development of relevant technologies. To address this issue, we proposes the first side-channel attack-specific dataset for the Dilithium—Dilithium Side-Channel Attacks Dataset (DSCAD), providing an experimental foundation for the application of deep learning in post-quantum cryptography research.Experimental environment：The DSCAD dataset was collected using the ChipWhisperer UFO platform, equipped with an STM32 F405 RGTx microprocessor running at 25 MHz and a 10 MHz passive crystal oscillator. Power consumption data was captured using a Pico 3206D oscilloscope with a 125 MHz sampling rate and an 8-bit mode. A BLP-48+ 50M low-pass filter was used to reduce noise. The Dilithium open-source reference implementation, set to NIST security level 2, was compiled using the gcc-arm cross-compiler with the -O1 optimization to simulate real-world scenarios.Target Operation：The dataset focuses on capturing the power traces during the storage of the sensitive polynomial $\mathbf{u}$ in the Dilithium implementation. Specifically, let \( u_{i,j} \), \( c_j \), and $(\mathbf{s}_1)_{i,j}$ be the coefficients of the polynomials $\mathbf{u}$ , $\mathbf{c}$ , and $(\mathbf{s}_1)$, where \( 1 \leq i \leq 4,1 \leq j \leq 256 \). They satisfy the following Equation (1):$$u_{i,j}=(c_j(\mathbf{s}_1)_{i,j}-c_j(\mathbf{s}_1)_{i,j}q^{-1})\mathrm{mod}^\pm2^{32}\cdot q)>>32(1)$$where \( q = 8380417 \) and $q^{-1}= 58728449$ are precomputed constants. The dataset includes 50,000 power traces, each with 40,000 samples, covering 1024 storage operations of $\mathbf{u}$. Out of these, 40,000 traces are used for signing random plaintexts with a random private key, while the remaining 10,000 traces are used to sign with a fixed private key to attempt key recovery.