Adversarial attacks pose a significant threat to machine learning systems, particularly when they exhibit transferability: the ability of a perturbation crafted for one "source" model to successfully fool an unknown "target" model. While current generative attacks attempt to maximize disruption at the layer level, they often suffer from a lack of diversity, focusing on a narrow set of visual concepts that may not be critical to the target model's decision-making process.
NAT (Neuron Attack for Transferability) introduces a paradigm shift by moving from layer-level disruption to targeting individual neurons. By training a diverse set of specialized generators, NAT achieves state-of-the-art transferability across a wide array of architectures, including standard CNNs, Vision Transformers, and hybrid models.
The Problem: Concept Collapse in Layer-Level Attacks
Existing generative adversarial attacks typically aim to maximize the distance between clean and adversarial embeddings at a specific middle layer of a network. However, researchers observed that these "layer-level" attacks often get stuck. They disproportionately disrupt only a few neurons representing a single visual concept (such as a specific texture), leaving the majority of the layer's neurons relatively unaffected.
If the target model does not rely heavily on that specific concept, the attack fails. To overcome this, an attack must be able to disrupt a wider variety of internal representations, creating diverse perturbations that can exploit different vulnerabilities in various target architectures.
The NAT Framework: A Mixture of Experts
NAT addresses the diversity problem by treating individual neurons as specific targets for disruption. The framework operates in several key stages:
- Neuron-Specific Generators: Instead of a single generator for an entire layer, NAT trains a separate generator ($G_j$) for each neuron $j$ in a target layer. Each generator is optimized to maximize the activation change for its specific neuron.
- Efficient Selection Phase: Training hundreds of generators is computationally prohibitive. NAT utilizes a "low-fidelity" training phase on a small subset (3.12%) of ImageNet to evaluate the transferability of all 512 neurons in a VGG16 source layer.
- Top-k Expert Training: The top $k=40$ neurons showing the highest transferability are selected. These "expert" generators are then fully trained on the complete ImageNet dataset.
- Diverse Inference: During an attack, an image can be processed by any of these 40 generators. Because each targets a different "concept" (e.g., vertical lines vs. color blobs), the resulting perturbations are highly diverse and complementary.
Experimental Results
The effectiveness of NAT was validated across a massive suite of 41 ImageNet-pretrained models, including 16 Convolutional, 7 Transformer, and 18 Hybrid architectures. It was also tested for cross-domain transferability on datasets like CUB-200-2011 (birds) and Stanford Cars.
Superior Transferability and Multi-Query Performance
In single-query settings, NAT outperformed leading baselines like LTP and BIA by over 14% in cross-model scenarios. Its advantage becomes even more pronounced in multi-query settings. Because the 40 generators are complementary, trying a few different perturbations leads to a significant jump in success rates.
Breaking Robust Architectures
One of the most impressive results is NAT's performance against Vision Transformers (ViTs). Despite being crafted on a standard VGG16 CNN, NAT achieved a 25% average fooling rate on ViTs, significantly higher than the ~18.9% achieved by previous state-of-the-art methods. This suggests that neuron-level attacks capture fundamental visual features that are shared across radically different model architectures.
Conclusion and Future Work
NAT demonstrates that the key to adversarial transferability lies in the diversity of the attack. By shifting the focus to individual neurons, the framework avoids conflicting objectives and ensures a broad disruption of the target model's internal representations. Future research may focus on developing "gating mechanisms" to automatically select the optimal generator for a specific input image and further optimizing the computational efficiency of training multiple experts.
Summary updated