Dynamic Gradient Balancing for Enhanced Adversarial Attacks on Multi-Task Models

Multi-task learning (MTL) creates a single machine learning model called multi-task model to simultaneously perform multiple tasks. Although the security of single task classifiers has been extensively studied, there are several critical security research questions for multi-task models including 1) How secure are multi-task models to single task adversarial machine learning attacks, 2) Can adversarial attacks be designed to attack multiple tasks simultaneously, and 3) Does task sharing and adversarial training increase multi-task model robustness to adversarial attacks? In this paper, we answer these questions through careful analysis and rigorous experimentation. First, we develop na\"ive adaptation of single-task white-box attacks and analyze their inherent drawbacks. We then propose a novel attack framework, Dynamic Gradient Balancing Attack (DGBA). Our framework poses the problem of attacking a multi-task model as an optimization problem based on averaged relative loss change, which can be solved by approximating the problem as an integer linear programming problem. Extensive evaluation on two popular MTL benchmarks, NYUv2 and Tiny-Taxonomy, demonstrates the effectiveness of DGBA compared to na\"ive multi-task attack baselines on both clean and adversarially trained multi-task models. The results also reveal a fundamental trade-off between improving task accuracy by sharing parameters across tasks and undermining model robustness due to increased attack transferability from parameter sharing.