Abstract
Deep Neural Networks (DNNs) are increasingly deployed in critical Cyber-Physical Systems (CPS), such as autonomous vehicles and medical devices. However, these models remain vulnerable to both software-based adversarial noise and hardware-based fault injections. This paper provides a comprehensive overview of these threat vectors, demonstrating that while adversarial attacks are potent, they are often fragile against standard image pre-processing. Conversely, hardware bit-flipping can collapse a model's accuracy with as few as 10 precise modifications. We evaluate these vulnerabilities using standard benchmarks, highlighting the need for robust defense mechanisms in real-world AI deployments.
Introduction
As DNNs transition from research environments to life-critical applications, their security becomes a paramount concern. This research addresses two primary vulnerabilities: adversarial attacks and fault-injection attacks. Adversarial attacks involve small, often imperceptible perturbations to input data—such as a specific sticker on a traffic sign—that lead to misclassification. Fault-injection attacks involve physical tampering with the hardware (e.g., via lasers or electromagnetic interference) to flip bits in memory or processing units, thereby corrupting the model's logic.
A critical observation made by the authors is that most adversarial research focuses on "naked" models, ignoring the pre-processing pipelines (like resizing or filtering) used in real-world systems. These pipelines can inadvertently neutralize attacks, creating a gap between theoretical vulnerability and practical exploitability.
Attack Methodologies and Datasets
The authors categorize attacks into several taxonomies:
- Adversarial Taxonomy: Distinguishes between Evasion (modifying inputs at inference time) and Poisoning (corrupting the training dataset). Evasion is further categorized into gradient-based (e.g., FGSM, C&W), score-based, and decision-based methods.
- Fault-Injection Techniques: Includes Bit-Flip Attacks (BFA), which utilize Row-Hammer or laser injection to target specific vulnerable bits in model weights, and Single Bias Attacks (SBA), which target output layer biases.
The effectiveness of these attacks was evaluated using the German Traffic Sign Detection Dataset (GTSRB) for autonomous driving scenarios and ImageNet for large-scale architectures like AlexNet and ResNet-50.
Key Results and Insights
The research yields two major findings regarding model robustness and vulnerability.
The Pre-processing Defense
A significant insight is the fragility of sophisticated adversarial attacks. When noise is injected at the sensor level, simple low-pass filters in the pre-processing stage—such as Local Average with neighborhood Pixels (LAP) or Local Average with Radius (LAR)—can effectively "blur away" the adversarial noise.
As shown in the experimental results, introducing these filters significantly boosts Top-5 accuracy. For instance, accuracy under a JSMA attack recovered from approximately 10% to nearly 89% with the addition of input pre-processing.
Extreme Vulnerability to Bit-Flips
In contrast to the potential robustness against adversarial noise, DNNs show a terrifying sensitivity to hardware-level bit-flips. Because weights are stored in floating-point or quantized formats, flipping a single bit in a critical location (like a weight's exponent) can cause catastrophic failure.
- AlexNet: Accuracy plummeted to nearly 0% after only 12 bit-flips out of 488 million bits.
- ResNet-50: Accuracy dropped to nearly 0% after only 10 bit-flips out of 204 million bits.
Conclusion and Future Work
The study concludes that the current state of DNN security is a "cat-and-mouse" game where defenses against one vector often leave others exposed. As computing moves to the "edge," physical access makes hardware-level fault injection a practical threat.
Future research should focus on developing filter-aware attacks, formal verification for safety regions, and hardware-level resilience such as range restriction to prevent single bit-flips from creating extreme weight values.
Summary updated