2026
Network-Level Prompt and Trait Leakage in Local Research Agents
USENIX Security 2026
TL;DR. Ironically, local web research agents leak surprisingly rich information through encrypted traffic metadata alone. We show passive observers can infer user prompts and persona traits, then evaluate lightweight mitigations that reduce leakage with minimal utility loss.
Bias Similarity Measurement: A Black-Box Audit of Fairness Across LLMs
ICLR 2026
TL;DR. Instead of asking whether one model is "fair," BSM, a fairness evaluation method, asks which models share the same bias patterns. This makes LLM fairness auditing more useful for comparing model families, releases, and procurement choices.
Preprint
AI Snitches Get Glitches: Towards Evading Agentic Surveillance
arXiv preprint, 2026
TL;DR. Agents with tool access brings additional surveillance risk. We formalize agentic surveillance, introduce SurveilBench, and show prompt-injection-style techniques can evade or disrupt surveillance agents.
Whose Agent Are You? Multi-Layer Fingerprinting and Attribution of Autonomous Web Agents
arXiv preprint, 2026
TL;DR. AI agents leave fingerprints across both network behavior and browser interaction patterns. This work shows a website owner can identify and attribute each agent in addition to distinguishing them from humans and traditional crawlers traffic.
2025
FedCC: Robust Federated Learning against Model Poisoning Attacks
Accepted to SecureComm 2025
TL;DR. FedCC defends federated learning against model poisoning attacks even when benign clients are non-IID. It clusters clients using representation similarity, filtering malicious updates without treating natural data diversity as an attack.
2022
Security and Privacy Issues and Solutions in Federated Learning for Digital Healthcare
FDSE 2022
TL;DR. Federated learning is appealing for healthcare, but its communication-heavy training loop creates new privacy and security risks. This work organizes FL attack surfaces, defenses, and research directions for digital health deployments.
Open problems in medical federated learning
IJWIS 2022
TL;DR. Medical FL needs more than a privacy-preserving training protocol. This paper turns real deployment barriers into concrete categories and guidelines for building medical federated learning systems.
2021
Federated Learning: Issues in Medical Application
FDSE 2021
TL;DR. This early medical FL work identifies the practical issues that make privacy-preserving collaboration hard in real settings. It focuses on heterogeneity, client management, traceability, security, and experimental framework design.
☠️
Are You a Good Client? Client Classification in Federated Learning
ICTC 2021
TL;DR. Not every unusual client in FL is malicious; some are just non-IID. This work separates anomalous and benign clients so robust aggregation does not punish legitimate data diversity.
Personalized federated learning with clustering: Non-iid heart rate variability data application
ICTC 2021
TL;DR. Medical data are rarely IID, so a single global FL model can underfit individual differences. This work uses clustered personalization to improve depression severity prediction from heart-rate variability data.
👁️
Visual Adversaries in AI Agent Web Navigation
UMass URV Showcase
TL;DR. Visual web cues can steer multimodal agents away from their intended tasks. This poster explores hidden or semi-transparent adversarial text and UI drift as security risks for agent web navigation.