Deep learning models for point cloud classification are highly vulnerable to adversarial attacks, while recent advances in diffusion-based purification have shown promising defensive performance. However, existing diffusion-based purification methods harbor two fundamental limitations. First, a distributional gap arises from their training on clean-to-clean paths, which fails to generalize to the required adversarial-to-clean transition. Second, a semantic mismatch occurs because the fixed victim classifier cannot adapt to the decision boundaries of the purified data distribution. To address this, we propose PANDA, a two-stage framework that combines robust purification with classifier adaptation. For purification, we introduce PANDA-P, a novel dual-branch diffusion training strategy that simultaneously optimizes on both clean-to-clean and adversarial-to-clean paths. This unified formulation boosts the purification effectiveness while preserving fidelity. For adaptation, we design PANDA-A, a fine-tuning scheme that leverages a consistency-driven learning objective to reshape the classifier’s feature space and recalibrate a robust decision boundary for the purified data. Extensive experiments show that PANDA achieves consistently superior robustness over existing purification-based defenses on both synthetic and real-world benchmarks.
Yu-Xin Zhang, Xiaofeng Cong, Minjing Dong et al.· IEEE Transactions on Informa...· 0 citations
Agent self-evolution in long-horizon LLM systems is largely procedural: useful experience is not merely stored information, but reusable procedures for searching, debugging, and verification. Yet current evaluations do not isolate this form of transfer. Agent benchmarks test single-episode task solving; memory benchmarks target information retention rather than procedural reuse. We introduce EvoAgentBench, a benchmark for agent self-evolution via Ability-guided transfer across four agentic domains: web research, algorithmic reasoning, software engineering, and knowledge work. EvoAgentBench extracts trace-grounded Abilities from agent executions, canonicalizes them into operational units, and builds domain-specific Ability Graphs linking tasks that share procedural overlap. By design, every test task is backed by verified training-side Ability support. Across a 528/267 train/test split, two scaffolds, and three backbones, curated Ability content transfers reliably across model families, but no current automatic method sustains positive gain in all settings. EvoAgentBench shifts self-evolution evaluation from aggregate accuracy comparison to fine-grained diagnosis of experience encoding, routing, and uptake. The benchmark is publicly available at https://huggingface.co/datasets/EverMind-AI/EvoAgentBench.
