Abstract
Adversarial attacks—subtle perturbations that trick Deep Neural Networks (DNNs) into misclassifying images—are a major hurdle for deploying AI in safety-critical fields. While many attack methods exist, they often lack fine-grained control over attack parameters. This paper introduces mFI-PSO, a white-box method that combines differential geometry with swarm intelligence to provide unprecedented flexibility. By utilizing a manifold-based influence measure to identify vulnerable pixels and Particle Swarm Optimization (PSO) to fine-tune perturbations, mFI-PSO allows users to specify the number of pixels changed, the desired misclassification probability, and the target class. Experimental results on MNIST and CIFAR-10 demonstrate that mFI-PSO outperforms several classic methods in both success rate and flexibility.
Introduction
Most existing adversarial attacks, such as FGSM or DeepFool, are relatively rigid. They typically find a perturbation that works but do not allow for precise constraints, such as changing a specific number of pixels or achieving a target confidence level for misclassification. Furthermore, traditional methods for identifying vulnerable pixels often rely on Euclidean-space measures like the Jacobian norm, which may ignore the intrinsic geometry of a model’s decision boundaries. There is a critical need for an adversarial framework that is both geometrically aware and highly customizable to better understand and defend against these vulnerabilities.
Methods
The mFI-PSO framework operates as a two-stage pipeline designed for precision engineering of adversarial examples.
1. Manifold-based First-order Influence (mFI)
Instead of relying on simple gradients, mFI-PSO employs a Riemannian manifold approach. The model's prediction probabilities are treated as a manifold, and mFI measures how a tiny change in a pixel "moves" the prediction on this manifold. This geometric approach identifies influential pixels and vulnerable images more accurately than standard saliency maps.
2. Particle Swarm Optimization (PSO)
Once vulnerable pixels are identified, the attack uses PSO—a gradient-free, population-based optimization algorithm—to determine the optimal perturbation values. Because PSO does not require a differentiable objective function, it enables the use of complex, user-defined loss functions. This allows for three main "customization knobs":
- $m$: The exact number of pixels to perturb.
- $P_{err}$: The specific confidence level of the resulting misclassification.
- $y_{target}$: The specific incorrect class the model should predict.
Key Insights and Ideas
A central finding of this research is that geometry matters. The model's internal representation is not a flat Euclidean space. By using mFI, the attack identifies pixels that are "intrinsically" important to the model's decision process.
As shown in the mFI maps, the vulnerability of specific pixels varies significantly depending on the target class ($y$). This suggests that mFI is an excellent tool for pre-screening datasets to identify images and pixels most likely to fail under specific adversarial conditions. Additionally, using PSO decouples the attack from mathematical constraints of standard loss functions, allowing for non-standard goals like hitting a specific probability target.
Key Results
The authors evaluated mFI-PSO on the MNIST and CIFAR-10 benchmarks using a ResNet32 architecture.
- High Success Rates: On vulnerable subsets, mFI-PSO achieved nearly 100% success rates, even with a restricted perturbation magnitude ($\epsilon$) of 0.15.
- Superiority over Classics: mFI-PSO significantly outperformed established methods like JSMA and DeepFool under constrained conditions. For instance, on a CIFAR-10 subset, mFI-PSO maintained a ~92% success rate while other methods' performance dropped significantly.
- Robustness Boost: Adversarial fine-tuning using mFI-PSO resulted in models that were more resilient to attacks than those trained with FGSM or PGD, indicating that mFI-PSO examples serve as more challenging training data.
Conclusion
While highly effective and flexible, mFI-PSO has limitations. As a population-based method, PSO is computationally more expensive than "single-shot" attacks like FGSM. Additionally, it currently requires white-box access to model parameters for mFI calculations. Future work will explore scaling this method to high-resolution datasets like ImageNet and optimizing the PSO search for greater efficiency in larger input spaces.
Summary updated