View Proposal

---

Proposer

John See

Title

Adversarial Robustness in Micro-Expression Recognition (TAKEN)

Goal

The main aim of the project is to establish a systematic robustness benchmark for micro-expression recognition (MER) models across different adversarial attacks and perturbation strengths, identify vulnerable facial regions that affect decisions, and to design an adversarial training strategy that improves robustness in a targeted manner.

Description

This project investigates the adversarial robustness of facial micro-expression recognition systems through a two-pronged approach. (1) A systematic benchmark will evaluate existing MER models under different adversarial attack types and strengths. Grad-CAM and related visualisation techniques will be used to examine which facial regions or action units are most susceptible to adversarial perturbations. Second, the project will investigate a targeted adversarial-training strategy based on the apex frame—the frame generally considered to contain the most discriminative information within a micro-expression sequence. Rather than applying uniform perturbations across all frames during adversarial training, stronger perturbations will be applied to the apex frame to examine whether this can improve model robustness while preserving recognition performance.

Resources

Publicly available micro-expression datasets (i.e. CASME II, SAMM, SMIC) are available for use. High-performance computing facilities (globally and in Malaysia) are available.

Background

Facial Micro-Expression Recognition (MER) aims to identify subtle and involuntary facial movements that occur for a very short duration. Although deep learning has achieved promising performance in MER, these models may be vulnerable to adversarial attacks, where carefully designed perturbations can cause incorrect predictions while remaining difficult for humans to perceive. Despite growing research on adversarial robustness in computer vision, the security and robustness of micro-expression recognition systems remain relatively underexplored.

Url

Difficulty Level

High

Ethical Approval

None

Number Of Students

1

Supervisor

John See

Keywords

adversarial robustness, micro-expression recognition, emotions

Degrees

Bachelor of Science in Computing Science