Mohamed Elmahallawy

SecureGate is a token-gated dual-adapter framework for federated LLMs that learns when to reveal PII. A lightweight gating module dynamically routes inputs through a privacy-preserving or utility adapter, selectively masking sensitive tokens before they influence model outputs while preserving task performance.

TKAR introduces a tokenized key-gating mechanism for LLM adapters that enforces fine-grained access control at the token level, preventing unauthorized extraction of private training data while maintaining model utility across federated clients.

CLIPGuard detects and purifies stealthy embedding-space backdoor attacks in CLIP. It partitions input images into grid cells, computes per-cell embedding sensitivity shifts to form a heatmap, isolates high-sensitivity regions via targeted segmentation, and inpaints them to remove triggers—achieving robust purification without model access.

AsyncFLEO addresses the fundamental privacy gap in asynchronous federated satellite learning: standard secure aggregation fails to protect long-term privacy when staggered orbital windows leak temporal update patterns. We propose differential privacy with adaptive noise calibration that provably preserves long-term privacy across LEO orbital windows.

We propose a multimodal vision-language framework for real-time situational awareness in underground mining disasters. The system combines dense visual reasoning over thermal and RGB imagery with language-grounded scene understanding to explain hazardous conditions invisible to traditional detectors, enabling timely emergency response.

OrbitChain is a blockchain-backed federated satellite learning framework enabling trustworthy multi-vendor collaboration in LEO networks. It offloads consensus to HAPs, ensures auditable model update provenance across orbits, and prevents manipulated contributions from corrupting global FL aggregation with sub-second finality.

FL-SEC empowers LEO satellites to execute large-scale ML tasks onboard using a personalized divide-and-conquer strategy. Each satellite trains a binary classifier for its local scene class, reducing communication rounds by 30× and energy consumption to 1.38 W with 96% accuracy. Awarded Best Paper Runner-up at IEEE PerCom 2024.

NomaFedHAP uses HAPs as distributed parameter servers and hybrid NOMA-OFDM waveforms to enable fast, bandwidth-efficient FL model transmissions across LEO constellations. We derive closed-form outage probability expressions and demonstrate significant convergence speed improvements over conventional FL-LEO methods.

CAV-AD distinguishes abnormal sensor readings from multiple simultaneous anomaly types while identifying malicious sensors in CAV networks. It achieves 98% average accuracy and 89% F1 score using an O-OS-CNN with dynamic amplification blocks that boost anomaly signals selectively.