Paper 2025/1728

Precision Strike: Targeted Misclassification of Accelerated CNNs with a Single Clock Glitch

Abstract

Fault injection attacks (FIAs) present a significant threat to the integrity of deep neural networks (DNNs), particularly in hardware-accelerated deployments on field-programmable gate arrays (FPGAs). These attacks intentionally introduce faults into the system, leading the DNN to generate incorrect outputs. This work presents the first successful targeted misclassification attack against a convolutional neural network (CNN) implemented on FPGA hardware, achieved by injecting a single clock glitch at the final layer (argmax) to manipulate the predicted output class. Our attack targets the commonly adopted argmax layer, a lightweight replacement for softmax in resource-constrained implementations. By precisely injecting a single clock glitch during the comparison phase of the argmax operation, the attack reliably induces misclassifications, forcing the model to 'skip' a specifically chosen class and output an incorrect label for it without affecting the computed scores of other classes. Unlike prior works that only cause random misclassifications, our attack achieves a high success rate of 80–87% for a targeted class, without inducing collateral misclassifications of other classes. Our evaluations show a significant reduction in classification accuracy, with the model’s performance dropping from an initial 94.7% to an average final accuracy ranging from 7.7–14.7%. Our attack is demonstrated on a CNN model implemented using a common systolic array architecture, which is well-suited for resource-constrained edge devices and artificial intelligence (AI) accelerators. Our study confirms the vulnerability of hardware-accelerated machine learning systems to low-cost physical attacks, emphasizing the critical need for hardware-level countermeasures in safety-critical machine learning applications.