GMML: Gradient-Modulated Robustness for Imbalance-Aware Multimodal Learning

摘要

Multimodal learning integrates diverse modalities to enhance robustness, yet real-world scenarios suffer from heterogeneous imbalance phenomena (noise interference, modality partial missing, intermodal information disparities), degrading performance through biased feature representations. Existing methods fail to adaptively modulate models under dynamic imbalance conditions. We propose GMML, a framework dynamically balancing multimodal gradients to counteract imbalance-induced biases: i) An imbalance-aware gradient modulation adaptively identifies contributions with smooth weight transitions to balance conflicting gradients; ii) A parameter constraint method enforces ℓ2-norm constraints on encoders, suppressing parameter oscillations and blocking noisy updates under modality missing/noise. Theoretically, GMML achieves a larger certified radius upper bound for complex imbalances, with convergence radius analysis providing theoretical guarantees. Experiments demonstrate superior robustness against three imbalance types, outperforming state-of-the-art by 3.3% and 2.3% in accuracy on KS and UCF-101 benchmarks. series Code: https://github.com/zhangzikai-security-ML/GMML.