Abstract
Recent works in adversarial robustness have proposed defenses to improve the robustness of a single model against the union of multiple perturbation types. However, these methods still suffer significant trade-offs between robustness to different attacks and performance on clean data. This paper introduces PROTECTOR, a novel framework that categorizes perturbation types before applying specialized robust predictors, effectively bypassing the "jack-of-all-trades" limitation of previous multi-robust models.
Introduction
In adversarial machine learning, achieving robustness against diverse attack types is a persistent challenge. A model trained to resist $L_\infty$ attacks (small, pervasive pixel changes) often remains vulnerable to $L_1$ attacks (large changes to a few pixels). Current "multi-robust" defenses attempt to train a single model to handle the union of these threats. However, this often leads to a "robustness trade-off," where the model performs worse on specific attacks than a dedicated specialist, and a "clean accuracy tax," where performance on unperturbed images degrades.
PROTECTOR (Perturbation Type Categorization for Robustness) addresses this by asking: Can we categorize perturbations first to avoid these trade-offs?
The PROTECTOR Framework
PROTECTOR utilizes a two-stage pipeline to handle adversarial inputs:
- Stage 1: The Perturbation Classifier ($C_{adv}$): Instead of identifying the object in the image, this model predicts the type of noise added to it. Theoretical analysis suggests that different perturbation types shift data into distinct, separable distributions.
- Stage 2: Specialized Robust Predictors ($M_A$): Once categorized, the input is routed to a specialist model trained specifically for that perturbation type.
Handling Adaptive Attackers
A critical component of PROTECTOR is its resilience to attackers who know the system's architecture. If an attacker tries to fool the Stage 1 classifier (e.g., making an $L_\infty$ attack look like an $L_1$ attack), they are forced to fundamentally alter the noise structure. This creates a natural tension: to successfully deceive the classifier, the attacker must often weaken the attack's effectiveness against the subsequent specialist. To further harden the system, PROTECTOR employs an ensemble of classifiers across pixel and frequency domains and utilizes small amounts of random noise to obscure decision boundaries.
Why Categorization Works
The core insight of this work is that adversarial attacks leave unique "fingerprints." While noise may appear random to the human eye, different mathematical constraints ($L_p$ norms) create statistically distinct patterns in the data distribution.
As shown in the feature space analysis, perturbations like $L_1$ and $L_\infty$ occupy different regions, making them highly categorizable. Furthermore, the authors found that the Fourier (frequency) domain is particularly effective for distinguishing these noise types. This separability allows PROTECTOR to achieve the high performance of a single-attack specialist while maintaining the broad coverage required for multi-attack defense.
Key Results
Evaluated on MNIST, CIFAR-10, and CIFAR-10-C, PROTECTOR consistently outperformed state-of-the-art multi-perturbation defenses such as MSD and MAX training:
- Enhanced Robustness: On CIFAR-10, PROTECTOR improved robustness against the union of $L_1, L_2,$ and $L_\infty$ attacks by 10% over previous methods.
- Preserved Clean Accuracy: PROTECTOR maintained a 7% higher benign accuracy on CIFAR-10 compared to competitors, effectively mitigating the "clean accuracy tax."
- Generalization: The perturbation classifier generalized remarkably well to unseen corruptions (e.g., frost or motion blur), categorizing them into broader groups with over 90% accuracy.
- Resilience: Even under full-knowledge adaptive attacks targeting both stages, PROTECTOR remained significantly more robust than traditional unified models.
Limitations and Future Work
While effective, PROTECTOR's performance is tied to the availability of high-quality specialist models for each attack type. Introducing a mathematically unique new attack would require training a new specialist. Additionally, the two-stage process introduces some inference latency, though the perturbation classifier itself is designed to be lightweight. Future research may explore more fine-grained attack sub-classes and the use of generative models to synthesize diverse perturbations for even more robust Stage 1 training.
Summary updated