Four AAAI‑23 Papers from Ant Security Lab on Adversarial 3D Point Clouds, GNN‑Based Anti‑Money Laundering, Spiking Neural Network Dynamic Graph Learning, and Differential‑Private Adaptive Clipping

Ant Security Lab continues its work on trustworthy AI, presenting four papers accepted at AAAI‑23 that address adversarial robustness, financial crime detection, efficient graph learning, and privacy‑preserving training.

PF‑Attack: Transferable 3D Adversarial Point Clouds

The paper proposes a simple yet effective method that jointly optimizes a perturbation and its randomly factorized sub‑perturbations to generate 3D point clouds that transfer across black‑box models. Experiments on ModelNet40 show 100% success in white‑box settings and superior transfer rates compared to baselines.

AMAP: GNN‑Based Iterative Risk Sub‑graph Extraction

AMAP introduces a graph neural network framework with an adaptive sub‑network extractor to detect money‑laundering transactions and simultaneously discover suspicious sub‑graphs. Evaluations on two real transaction datasets demonstrate consistent improvements in ROC, AUC‑PR, and G‑mean over existing methods.

SpikeNet: Efficient Dynamic Graph Learning via Spiking Neural Networks

SpikeNet leverages leaky‑integrate‑and‑fire neurons to encode graph snapshots as sparse spike sequences, achieving competitive node‑classification accuracy while reducing parameter count and training time. Experiments on three large‑scale dynamic graph benchmarks confirm its superiority in both performance and energy efficiency.

DP‑PSAC: Per‑Sample Adaptive Clipping for Differential Privacy

DP‑PSAC replaces constant clipping in DP‑SGD with a non‑monotonic weight function that adapts to gradient magnitude, eliminating the need to tune the clipping bound C . Theoretical analysis and experiments on image and text classification tasks show state‑of‑the‑art accuracy across multiple datasets.

Expert commentaries from professors at Zhejiang University and Sun Yat‑sen University highlight the practical impact and research significance of these contributions.