Security, Privacy, and Forensics Investigation of Multimodal Human-AI Interaction

🧠🤝🔒 Research Description

🔍 Overview
As human-AI interaction becomes increasingly multimodal, encompassing text, voice, image, video, gesture, and immersive interfaces, the boundaries between personal space, behavioral data, and AI systems are dissolving. This long-term research agenda investigates the security, privacy, and forensic implications of these evolving interfaces, aiming to establish a rigorous foundation for safeguarding users and holding AI systems accountable across modalities and platforms.

🧪 Research Thrusts
Our vision integrates three interrelated thrusts:

Security and Privacy in Human-AI Collaboration: We investigate vulnerabilities in collaborative and interactive environments, such as virtual conferencing platforms, shared whiteboards, and mixed-reality spaces, where AI acts as both an enabler and observer of user interaction. Our work targets real-time communication streams, digital co-creation tools, and behavioral sensing features, applying privacy-preserving algorithms (e.g., local differential privacy, on-device obfuscation) to protect user identity, intent, and context without compromising collaboration efficiency.
Behavioral Privacy and Latent Inference Risks: In this thrust, we formalize a threat model for latent inference: the unauthorized extraction of personal attributes (e.g., age, gender, health status, emotional state) from non-semantic user behaviors, such as drawing patterns, voice modulation, or gaze direction, by multimodal AI models. We develop transformation techniques and PETs (privacy-enhancing technologies) to neutralize these inference pathways while retaining core task utility, laying the groundwork for privacy-aware interaction design in future AI systems.
AI Forensics for Multimodal Systems: We pioneer forensic investigation techniques to analyze, attribute, and audit behaviors in generative and assistive AI systems, such as large language models (LLMs), image/video generators, and embodied agents (e.g., AR/VR headsets, smart glasses, mobile devices). Our research develops tools and methodologies to trace model outputs, detect synthetic content, and reveal latent inference pathways. These capabilities are vital for auditing misuse, disinformation, deepfake attacks, and inadvertent privacy breaches.

🌐 Research Impact
This unified effort establishes a novel research discipline at the intersection of AI security, digital forensics, and human-computer interaction. It responds to urgent global needs: understanding AI misuse, protecting user privacy, and designing resilient collaborative systems in an era where AI models are not only tools, but actors with inference and surveillance capabilities. Our outcomes will shape next-generation AI regulation, human-centered AI design, and trustworthy multimodal interaction.

Current Projects

Security and Privacy of Human-AI Interaction

This research project addresses the growing need to understand and mitigate security and privacy risks in increasingly complex human-AI interaction modalities. As AI systems integrate into daily life through multimodal interfaces, including text, speech, handwriting, sketching, gesture, sensor data, and visual inputs, they capture rich behavioral and contextual information that extends far beyond explicit user intent. These signals, while enabling intuitive and powerful user experiences, also expose individuals to novel threats ranging from behavioral surveillance and identity inference to adversarial manipulation and data leakage. This project aims to build a comprehensive framework for analyzing, modeling, and securing the full spectrum of human-AI interactions. Our objectives are threefold:

Systematically investigate how different interaction modalities introduce privacy vulnerabilities and security risks;
Develop formal threat models that capture both adversarial and incidental inference capabilities of advanced AI systems, including large language models (LLMs), vision-language models (VLMs), and multimodal agents; and
Design and evaluate technical safeguards that preserve user privacy and system integrity without degrading utility.

Threat Modeling for LLMs and VLMs Adversarial Privacy in AI Interfaces Privacy-Preserving Technologies for AI User-Centric AI Security Behavioral Inference Risks in AI Systems

Multimodal AI Forensics

This study presents a holistic forensic analysis of AI-powered desktop and mobile applications, focusing on identifying and recovering digital artifacts for investigative purposes. LeveThis research thrust focuses on the development of forensic analysis techniques for emerging AI-powered systems that operate across text, image, audio, video, and immersive modalities. As generative models become embedded in everyday applications, wearables, and virtual environments, they generate complex digital traces that pose new challenges for accountability, evidence preservation, and privacy auditing. Our work aims to systematically recover and interpret artifacts produced by these systems to support post hoc investigations, security assessments, and compliance analysis. The objectives of this research are threefold:

Develop modality-aware forensic techniques capable of recovering and reconstructing user-AI interaction artifacts across platforms and interfaces;
Characterize the evidentiary value, security implications, and privacy risks of these artifacts through rigorous analysis; and
Design tool-supported methodologies that enable reproducible, scalable, and policy-aligned forensic investigations of generative AI systems.

Generative AI Applications: Investigation of desktop and mobile platforms to recover cached content, interaction logs, memory artifacts, and network traffic, enabling the reconstruction of user-AI exchanges and assessment of data governance practices.
Smart Wearables (e.g., Meta Glasses): Forensic recovery of multimedia captures, sensor data, and cloud-synced artifacts, with a focus on deletion resistance, metadata analysis, and risks to user and bystander privacy.
Virtual Reality Environments: Extraction and analysis of behavioral and content-level artifacts from immersive platforms, emphasizing the forensic traceability of AI-generated scenes and user interaction histories in environments powered by LLMs and VLMs.
Generative Edit Forensics: Development of techniques to detect, attribute, and verify AI-generated or manipulated content across modalities, supporting integrity verification, tampering detection, and evidentiary reliability in digital ecosystems.

Multimodal AI Forensics Generative AI Artifact Recovery Forensic Analysis of LLMs and VLMs Generative Edit Detection Human-AI Interaction Analysis

Publications

Mohd. Farhan Israk Soumik, Syed Hasan, Abdur R. Shahid, "Evaluating Apple Intelligence's Writing Tools for Privacy Against Large Language Model-Based Inference Attacks: Insights from Early Datasets." arXiv preprint arXiv:2506.03870 (2025)
Kankanamge, Malithi Wanniarachchi, Nick McKenna, Santiago Carmona, Syed Mhamudul Hasan, Abdur R. Shahid, and Ahmed Imteaj. "Digital Forensic Investigation of the ChatGPT Windows Application." arXiv preprint arXiv:2505.23938 (2025).
Mohd. Farhan Israk Soumik, W.K.M Mithsara, Abdur Rahman Bin Shahid, Ahmed Imteaj, Audio Editing Features As User-Centric Privacy Defenses Against Large Language Model (LLM)-Based Emotion Inference Attacks, The Sixth AAAI-25 Workshop on Privacy-Preserving Artificial Intelligence (AAAI PPAI 2025).
Abdur R. Shahid and Ahmed Imteaj, "Securing User Privacy in Cloud-Based Whiteboard Services Against Health Attribute Inference Attacks", In IEEE Transactions on Artificial Intelligence (2024).
Abdur R. Shahid, and Sajedul Talukder, "Evaluating Machine Learning Models for Handwriting Recognition-based Systems under Local Differential Privacy", 2021 Innovations in Intelligent Systems and Applications Conference (ASYU), pp. 1-6. IEEE, 2021

Page updated

Google Sites

Report abuse