Nominated computer vision tasks have nowadays performed on super-human level with the advances of deep learning techniques. However, adversarial machine learning researchers demonstrate that such vision systems could not yet be as robust as human systems. Acting as a new gamut of technologies, adversarial machine learning covers the research and development of studying the intact capabilities and malicious behaviors of machine learning models in an adversarial scenario. The potential vulnerabilities of ML models to malicious attacks can result in severe consequences for safety-critical systems. One most known manner is via imperceptible perturbations to the input images/videos. Although it is not to be alarmist, researchers in machine learning and computer vision areas have a responsibility to preempt attacks and build safeguards especially when the task is critical for information security, and human lives (e.g., autonomous driving systems). We need to deepen our understanding of machine learning in adversarial environments.

Thus motivated, we organize the ACCV 2022 workshop on “Adversarial Machine Learning towards Advanced Vision Systems” in Dec 2022, in conjunction with Asia Conferences on Computer Vision (ACCV 2022) in Macau, China. We will invite world-renowned keynote speakers to share their latest research progress in this direction at the workshop. Besides, we will invite researchers to submit their latest research output to this workshop and present them in a poster session.

Dec. 4th, 2022 (Beijing 9:00 - 12:00, GMT+8)

Session One (Beijing time 9:00 - 10:25):

9:00 - 9:45 Keynote Talk (Dr. Pin-Yu Chen, IBM Thomas J. Watson Research Center, USA)

Talk Title: Reprogramming Foundation Models with Limited Resources

Talk abstract: TBA

9:45 - 10:25 Keynote Talk (A/Prof. Dongxiao Zhu, Wayne State University, USA)

Talk Title: Empowering Explainable Machine Learning through the Lens of Adversarial Robustness and Fairness

Talk abstract: Deep Neural Networks (DNNs) are complex nonlinear functions parameterized by weights to map inputs to outputs. It has attracted much attention in machine learning community due to its state-of-the-art performance on various tasks. Despite the success, explainability of a complex DNN remain an open problem, hindering its wide deployment in safety and security-critical domains. In this talk, I will introduce our recent works on ensuring DNN explainability from the lens of adversarial robustness and fairness using both image/text classification and recommender systems. Regarding adversarial robustness, I will describe Adversarial Gradient Integration for explaining convolutional neural network (CNN) based image classification and detection, which utilizes adversarial examples for better estimating the input attributions (IJCAI-21). To incorporate the unique self-attention mechanism from Transformers in rendering explanations, I will continue with Attentive Class-Activation Tokens for explaining Transformer based text classification (NeurIPS-22). Regarding fairness, I will describe Counterfactual Interpolation Augmentation to enhance the explainability of CNN via fairness-aware data augmentation (IJCAI-22). I will conclude this talk with samples of other trustworthy AI systems developed in my research group, including an interpretable recommender system (IJCAI-20), an adversarial robust image classification system (AAAI-23, AAAI-21), and a driver-centered and resource-aware Electric Vehicle (EV) Charging recommender (ECML-22).

Session two (Beijing time 10:40 - 11:55):

10:40 - 11:05 Lili Zhang, Xiaodong Wang, ADVFilter: Adversarial Example Generated by Perturbing Optical Path

11:05 - 11:30 Qingguo Zhou, Ming Lei, Peng Zhi, Rui Zhao, Jun Shen and Binbin Yong, Towards Improving the Anti-attack Capability of the RangeNet++

11:30 - 11:55 Yanli Li, Abubakar Sadiq Sani, Dong Yuan and Wei Bao, Enhancing Federated Learning Robustness Through Non-IID Features

Dr. Minhui (Jason) Xue, Data61, CSIRO, Australia

Dr. Huaming Chen, The University of Sydney, Australia