Introduces a 360M-parameter Agentic Physical AI model whose control policies are validated through closed-loop reactor simulation. Scaling to 100K scenarios yields ~500× variance collapse and 92% success within ±1% target power.

Introduces LAMDA, a longitudinal Android malware benchmark with over 1 million apps spanning 12 years and 1,380 malware families. It exposes how concept drift degrades standard detectors and enables systematic study of temporal robustness.

Proposes a privacy-preserving system that detects and disrupts scams during active conversations while federated learning updates models without sharing raw data. It maintains ~0.80 engagement with PII leakage at or below 0.0085.

Introduces AgentFence, a 14-class trust-boundary framework spanning planning, memory, retrieval, tool use, and delegation. With the base model fixed, security-break rates vary from 0.29 to 0.51 across eight agent architectures, revealing architecture-dependent risk.

Reveals a safety blind spot in tool-using LLMs: across 768 evaluations, explicit schema violations occur in only 0.78% of cases while unsafe semantic intent appears in 65.95%, showing that schema compliance does not imply safe behavior.

Shows that benign domain fine-tuning alone can destroy guard-model safety geometry, driving Granite Guardian's refusal rate from 85% to 0%. The proposed FW-SSR regularization restores 75% refusal and reduces WildGuard attack success to 3.6%.

Introduces SIF, where individually benign subtasks compose into a policy-violating plan, succeeding in 71% of 14 enterprise scenarios while every subtask passes safety filters. Plan-level checking detects all observed attacks with 0% false positives.

Combines LLM-guided augmentation and reasoning with Dynamic UMAP and Bayesian personalized federated learning, aligning language-model representations with client-specific graph structure for privacy-constrained node classification and link prediction in low-resource settings.

Introduces CPRS, a unified privacy-risk score combining profile attributes, graph structure, and user-generated content. Evaluation on Facebook and Koo data identifies graph structure as the strongest average risk source, while 85% of users found the resulting dashboard clear and actionable.

Benchmarks LLMs for Bengali hate-speech detection across more than 120K samples, including transliterated and code-mixed text. Fine-tuned LLMs exceed 90% macro-F1, with a DeepSeek–XLM-R ensemble reaching 91.2%.

Combines RoBERTa-based emotion prediction with recursive comment-thread aggregation to track Approval, Toxicity, Hate, Threat, and related sentiment dynamics over time. The emotion model achieves AUC 0.92, with XGBoost attaining macro-F1 above 0.40.

Introduces dynamic proximity-based clustering and hybrid decentralized aggregation to reduce federated-learning communication costs. Across five datasets, SAFARI reduces communication requirements by up to 10× while accelerating training and improving energy efficiency.

Proposes distributed threat intelligence that combines lightweight edge models, LLM-guided adaptation, and collaborative knowledge sharing. Only suspicious network or system-log information is escalated, supporting responsive and privacy-preserving threat detection and device isolation.

Introduces dynamic clustering and hybrid decentralized aggregation to reduce federated learning's dependence on centralized infrastructure. SCALE achieves nearly 10× lower communication overhead while reducing training latency and energy consumption without sacrificing learning performance.

Develops an AI-enabled Zero-Trust architecture for nuclear power plants that combines continuous authentication with behavioral analytics, endpoint protection, network segmentation, and continuous monitoring to prevent malware without assuming implicit trust in users or devices.

Combines zero-knowledge proofs with Hyperledger Fabric to verify social-network identities without exposing sensitive identity information. In a 50-user prototype, the multilayer framework prevented fraudulent, cloned, and multiple-account creation with 100% success.

Tailors federated learning to collaborative intrusion detection across connected autonomous vehicles, allowing vehicles to strengthen collective detection without centralizing local data. The framework examines attack and defense mechanisms for privacy-preserving, dependable CAV security.

Combines BERT-based emotion scoring with bottom-up aggregation of comment threads to monitor evolving social-media sentiment. The emotion model achieves AUC 0.91, while Decision Tree models attain macro-F1 above 0.40.

Introduces an interface and classifier that nudges users to inspect suspicious Facebook friend requests. It reduces accepted sockpuppet invitations by 42.6 percentage points and predicts user decisions with an F1 score of 96.52%.

Using epidemiological data from 24 countries, identifies a practical reopening signal: two weeks after infection rates cross below recovery rates while new cases continue declining. The finding is distilled into an interpretable large-scale reopening index for resource-constrained settings.

Extends AbuSniff into a large study of user-perceived friend abuse and its prediction from social activity. Across 263 participants, users accepted 78% of predicted defensive actions, while questionnaire responses were predicted with F-measures up to 89.7%.

Introduces AbuSniff to identify Facebook friends perceived as abusive or strangers and recommend defenses such as restricting, unfollowing, or unfriending them. Across 263 participants, users accepted 78.2% of automatically suggested actions without completing the abuse questionnaire.

Through two user studies involving 80 participants, establishes that Facebook friend relationships themselves can create privacy and content-abuse risks, with many users identifying at least one friend as potentially abusive or effectively a stranger.