Abstract
Resistive Random Access Memory (RRAM)-based Neuromorphic Computing Systems (NCS) have gained significant attention for their superior data processing capabilities and energy efficiency compared to traditional architectures. As these systems are increasingly deployed in data-centric applications, their reliability and security become critical concerns. This paper systematically investigates adversarial threats to RRAM-based NCS, revealing that inherent hardware variations can be leveraged to amplify attack effectiveness—a factor largely overlooked by traditional algorithmic attacks. We propose two hardware-aware attack methods: VADER (Variation-Oriented Adversarial Attack), which perturbs input samples to mislead predictions, and EFI (Enhanced Fault Injection), which modifies the parameter space to misclassify specific samples while maintaining overall accuracy. Experimental results demonstrate that these methods achieve nearly 100% success rates with minimal operational cost and high stealthiness.
Introduction
Neuromorphic Computing Systems (NCS) utilizing RRAM crossbars offer a promising path toward high-efficiency artificial intelligence by performing matrix-vector multiplications directly in memory. However, RRAM is a nanoscale technology subject to significant non-ideal properties, including manufacturing defects, resistance fluctuations, and random noise.
While the research community has traditionally focused on mitigating these variations to ensure reliability, this study explores their potential for exploitation. Traditional adversarial attacks and fault injection methods are typically "hardware-unconscious," treating the neural network as a mathematical abstraction. By ignoring the physical state of the hardware, these attacks are often less efficient and more susceptible to standard defenses. We demonstrate that by accounting for the specific noise profiles of RRAM hardware, attackers can significantly lower the barrier to compromising these systems.
Hardware-Aware Attack Methodologies
1. VADER: Variation-Oriented Adversarial Attack
The goal of VADER is to perturb input samples (e.g., images) to cause misclassification on a specific hardware instance. Unlike standard attacks like the Projected Gradient Descent (PGD), which rely on the model's global gradients, VADER introduces the concept of a gradient difference ($g_d$). This metric represents the discrepancy between an "ideal" software model and the "variation-affected" hardware model.
By identifying "variation-sensitive" pixels—those most likely to amplify the hardware's existing errors—VADER can push the system's inherent noise over the decision threshold. This allows the attack to succeed even against models hardened with software-level adversarial training, as the attack exploits the unique physical signature of the chip rather than just the model's mathematical boundaries.
2. EFI: Enhanced Fault Injection
EFI targets the model's weight parameters (synapses) to force a specific target misclassification for a chosen input. The innovation of EFI lies in its greedy victim parameter selection algorithm. Instead of fighting against hardware variations, EFI "cooperates" with them. It identifies RRAM cells that are already deviating toward the desired error state due to manufacturing noise. By selecting a minimal set of additional cells to modify, EFI achieves the attack with a fraction of the effort required by traditional methods.
Experimental Evaluation
Setup
The proposed attacks were evaluated using standard computer vision benchmarks:
- MNIST: 70,000 handwritten digits.
- CIFAR-10: 60,000 color images across 10 classes.
- Architectures: LeNet (for MNIST) and Wide Residual Networks (WRN for CIFAR-10), both quantized to 8-bit precision to reflect RRAM hardware constraints.
Key Results
- Attack Success Rate: Both VADER and EFI achieved nearly 100% success rates across the tested datasets.
- Stealth and Efficiency: VADER's perturbations remain human-imperceptible. Meanwhile, EFI demonstrated extreme efficiency; on CIFAR-10, it could force a targeted misclassification by modifying as few as 20 parameters in the final layers. This represents a >95% reduction in operational cost compared to traditional fault injection methods like GDA or Fault Sneaking.
- Defense Resilience: VADER successfully bypassed state-of-the-art adversarial training, highlighting a "security gap" where software-only defenses fail to account for hardware-level vulnerabilities.
Conclusion and Future Work
This research demonstrates that the physical imperfections of RRAM-based hardware can be weaponized to create highly effective and stealthy attacks. To counter these threats, future security strategies must be hardware-aware. We suggest two primary defense directions:
- On-device Adversarial Training: Incorporating the specific variation profile of the hardware into the training loop to close the security gap.
- Fault Detection Routines: Implementing periodic checks to identify malicious weight tampering that exceeds expected natural drift.
By acknowledging the interplay between software algorithms and hardware realities, we can develop more robust and secure neuromorphic systems.
Summary updated