New geometry-aware framework makes deep nets robust to adversarial attacks
Layer-wise local training creates smooth manifolds resistant to perturbations.
A new paper from researchers at an unnamed institution (likely Chinese, given author names) reveals a geometric explanation for adversarial vulnerability in deep neural networks and proposes a practical fix. The authors—Yixiong Ren, Wenkang Du, Jianhui Zhou, and Haiping Huang—introduce a geometry-aware learning framework that trains networks layer by layer, each layer locally optimizing its representation geometry. This process encourages features from the same class to cluster tightly while pushing different classes far apart, creating a smooth feature manifold that is inherently resistant to small perturbations.
The framework's performance is theoretically grounded in data-dependent statistical mechanics, where network parameters are integrated out, and a phenomenological model with Hebbian coupling between hidden representation elements. Crucially, the method allows the network to learn new information without catastrophic forgetting by reducing representation interference between old and new data. In experiments, models trained this way showed strong robustness against both white-box attacks (where the attacker knows the model) and black-box attacks (where the attacker doesn't), without sacrificing clean accuracy—a longstanding trade-off in adversarial defense.
- Layer-wise local training enforces intra-class compactness and inter-class separation in feature space.
- The framework uses statistical mechanics to derive robustness guarantees by integrating out network parameters.
- Achieves robust defense against both white-box and black-box attacks without compromising standard accuracy.
Why It Matters
Enables safer deployment of deep learning in security-critical areas like autonomous driving and medical imaging.