Comparative analysis of explainable-AI methods (Grad-CAM, Integrated Gradients, LIME) for lung-cancer classification on deep CNNs, quantifying computational-efficiency and interpretability tradeoffs for clinical diagnostic workflows.

Adversarial attack and defense framework (PGD, FGSM) against CNN architectures for dermatoscopic classification, achieving a 27.73 percentage-point robustness improvement through PGD-based adversarial training.

Hybrid intrusion detection combining CNN feature extraction and LSTM sequence modeling with an SVM classifier — 97.29% accuracy on multi-class network attack classification, outperforming traditional approaches by 15+ percentage points.

Swin Transformer-based image captioning with hierarchical shifted-window attention and geometric-aware self-attention, evaluated on MSCOCO and Flickr30k with BLEU, METEOR, and CIDEr.