SECURE-XED: A Unified Explainable CNN Framework with Hybrid Metaheuristic Optimization for Cross-Domain Malware and Deepfake Detection

We present SECURE-XED, a unified, explainable, and adversarial-aware learning system for two traditionally disjoint domains: Android malware classification and deepfake detection. The system uses a shared convolutional neural network backbone with domain-specific input adaptation and two task-specific heads. It integrates SHapley Additive exPlanations for explanation, the Fast Gradient Sign Method and Projected Gradient Descent for adversarial evaluation under Structural Similarity Index Measure guardrails, and a selective-activation controller that enables behavioral malware features when confidence is low or evasion is suspected. A hybrid Seagull Optimization Algorithm–Imperialist Competitive Algorithm procedure jointly tunes clean accuracy, robustness under projected-gradient attacks, expected calibration error, parameter count, and inference latency. SECURE-XED is evaluated on standard Android malware corpora and the FaceForensics++, Celeb-DF, and Deepfake Detection Challenge benchmarks. Under Projected Gradient Descent, Android malware accuracy decreases by approximately 19–21% with a perturbation budget of 0.1 over 40 steps, while deepfake accuracy decreases by 16.6–17.6% with a perturbation budget of 0.03 over 10 steps, demonstrating domain-dependent adversarial sensitivity. Under clean deepfake inference conditions, excluding post-prediction explanation generation and adversarial processing, the unified configuration reduces average inference latency from 34.1 to 26.9 milliseconds per frame and approximately halves the parameter requirement relative to maintaining separate model instances, while retaining equal or slightly better clean accuracy. A classroom-oriented simulator exposes explanation overlays and adversarial controls on real frames. A six-participant human-grounded pilot descriptively showed a 22-percentage-point change in Region Identification Accuracy, a 12-second reduction in decision time, and a 0.9-point Likert change in perceived clarity and trust; these pilot outcomes are descriptive and not inferential. The resulting framework combines shared cross-domain representation, multi-objective optimization, adversarial evaluation, explainability, and model-in-the-loop educational interaction within a single modular architecture.