Abstract
Class Activation Mapping (CAM) methods are widely used to interpret the decisions of deep neural networks due to their computational efficiency and intuitive visualizations. Although recent variants improve spatial granularity and localization accuracy, most do not explicitly optimize for faithfulness-the degree to which an explanation reflects the model's actual decision-making. In this paper, we propose DFO-CAM (Dual-Flow Optimization for Class Activation Mapping), a novel explainability framework that preserves the desirable properties of CAM while actively discovering class-relevant features that positively contribute to model predictions. DFO-CAM introduces a dual-flow optimization framework that jointly promotes informative, fine-grained evidence in the explanation heatmap and suppresses spurious activations, yielding a trade-off between visual comprehensibility and decision faithfulness. The framework is compatible with existing model architectures and CAM variants and can be used as a plug-in to enhance their faithfulness while preserving the characteristics of the original CAM methods. Experiments show that DFO-CAM produces more faithful explanations than strong base-lines while passing sanity checks and preserving class sensitivity. In addition, we conduct ablation studies to validate the effectiveness of the structural design. The code is available at https://anonymous.4open.science/r/AnonymousFor3AI-CQUPT1811/.