The black-box nature of deep neural networks poses a significant challenge to their deployment in high-stakes decision-making scenarios, conflicting with the principles of transparency and accountability in Responsible AI. To address this, this paper proposes HF-CAM (Hard Prototype Anchored Concept Activation Mapping), a novel post-hoc interpretability method. HF-CAM first employs K-Means clustering in the high-dimensional feature space to obtain hard prototype anchors with strong class discriminability. A parameterized Gaussian kernel function is then introduced to fuzzify the discrete hard partitions into smooth fuzzy membership maps, generating spatially localized concept activation maps. Systematic experiments are conducted on both Fashion-MNIST and CIFAR-10 datasets, with comparisons against Fuzzy C-Means based CAM (FCAM), Grad-CAM, Score-CAM, and a simplified ACE baseline. Quantitative results over multiple random trials show that HF-CAM consistently reduces class-prototype association entropy and improves concept purity across various prototype numbers. On CIFAR-10 with ResNet-18 at \(K=10\), HF-CAM reduces average entropy from 2.303 to 0.515 and improves concept purity from 0.103 to 0.758 (weighted mode); on Fashion-MNIST, it reduces entropy from 2.239 to 2.055 and improves purity from 0.146 to 0.206. Ablation studies on Gaussian bandwidth and sample size demonstrate the influence of the bandwidth parameter and the robustness of HF-CAM to different sample sizes. Counterfactual intervention experiments with random controls and paired t-tests indicate that the prototypes identified by HF-CAM have significantly stronger influence on model predictions than random masks or random prototypes, supporting their relevance to the decision process. HF-CAM provides relatively clear and referenceable visual explanations for understanding convolutional neural networks.