News
AI4CCAM workshop on data poisoning attacks
On 29 September, at Institut Mines Télécom (IMT) – Télécom SudParis, AI4CCAM organized a workshop on data poisoning attacks.
The widespread adoption of 3D point-cloud deep learning has greatly improved Connected and Autonomous Vehicles’ (CAVs) ability to perceive, classify, and react to road scenes. Validation of these systems relies on large simulated environments built from massive datasets. However, there is a shortage of such datasets, which is the reason practitioners commonly resort to data augmentation techniques such as Generative Adversarial Networks (GAN) to expand training corpora.
AI4CCAM believes that this reliance on shared datasets and augmentation pipelines creates a critical attack surface, where a malicious actor who introduces poisoned samples into the dataset ecosystem can have their influence amplified by augmentation, producing highly compromised scenarios and degraded downstream behavior.
The workshop had two primary objectives through practical sessions:
– Experimentally evaluate whether common augmentation techniques exacerbate poisoning attacks on 3D point-cloud data.
– Quantify the impact of poisoning attacks on CAV perception and downstream decision-making.
