Influence of Parallelism in Vector-Multiplication Units on Correlation Power Analysis

The use of neural networks in edge devices is increasing, which introduces new security challenges related to the neural networks’ confidentiality. As edge devices often offer physical access, attacks targeting the hardware, such as side-channel analysis, must be considered. To enhance the performance of neural network inference, hardware accelerators are commonly employed. This work investigates the influence of parallel processing within such accelerators on correlation-based side-channel attacks that exploit power consumption. The focus is on neurons that are part of the same fully-connected layer, which run parallel and simultaneously process the same input value. The theoretical impact of concurrent multiply-and-accumulate operations on overall power consumption is evaluated, as well as the success rate of correlation power analysis. Based on the observed behavior, equations are derived that describe how the correlation decreases with increasing levels of parallelism. The applicability of these equations is validated using a vector-multiplication unit implemented on an FPGA.

💡 Research Summary

**
The paper investigates how parallelism in vector‑multiplication units, which are the core of fully‑connected layers in neural‑network accelerators, influences the effectiveness of Correlation Power Analysis (CPA) attacks. The authors first decompose the total power consumption of a processing element (PE) into data‑dependent, operation‑dependent, constant, and noise components. For the purpose of theoretical analysis they assume the constant and noise terms are zero, thereby focusing on the exploitable part of the power that originates from the multiply‑and‑accumulate (MAC) operation and the data being processed.

In a typical accelerator each PE computes one MAC per clock cycle: it multiplies a weight (w_i) with the shared input value (x) and adds the result to an accumulator stored in a register. When (N) neurons of the same layer are processed in parallel, the overall power trace is the sum of (N) independent data‑dependent contributions. CPA attacks rely on building hypothetical intermediate values (usually the product (x\cdot w_i)) and converting them into predicted power values using a power model such as Hamming Weight (HW) or Hamming Distance (HD). The Pearson correlation coefficient between the predicted and measured power traces indicates whether a weight hypothesis is correct.

The authors derive a simple yet powerful relationship for the correlation coefficient of the correct weight hypothesis as a function of the degree of parallelism:

\