Secure & Private Intelligent Systems Lab

Multidisciplinary Pathways to Cybersecurity Professionals

Welcome to Secure & Private Intelligent Systems Lab!

The Secure & Private Intelligent Systems (SPIS) Laboratory at Miami University was founded by Dr. Honglu Jiang in Fall 2022. Our research focuses on data and model privacy, as well as the security and robustness of machine learning, deep learning, and large language models (LLMs). We explore topics such as differential privacy, federated learning, Trustworthy AI for healthcare applications, next-generation Internet architectures, and cyber-physical systems. Our work has been published in leading journals and conferences, including IEEE Transactions on Mobile Computing, IEEE Transactions on Vehicular Technology, IEEE Transactions on Knowledge and Data Engineering, Computers & Security, IEEE ICASSP, and IEEE MASS. This lab's research has been supported by the National Institute of Standards and Technology (NIST).

Latest: Open call for Ph.D. and Master students. Feel free to drop me an email titled "Prospective PhD student: First_name Last_name" and attach your CV.

Featured Research

Membership Inference Attacks on Fine-Tuned Large Language Models

Membership Inference Attack (MIA) aims to determine whether a given sample was part of a target model's training data. Most existing MIAs against fine-tuned large language models rely on strong assumptions. We propose a MIA method, ECT-MIA, which exploits the memory of LLMs for entities and their contextual texts. We guide a reference model to learn from the samples to be detected, and then distinguish members from non-members by comparing the deviations in outputs between the target model and the reference model, specifically on entities and entity-related texts within the target samples. Our method operates as a purely black-box attack, requiring only access to the plain text outputs of the target model—without relying on internal parameters or any data with a distribution similar to the target model's training dataset.

Curse to Blessing: Leveraging Model Inversion Attacks to Detect Backdoor Attacks

We propose a novel approach that transforms model inversion attacks from a privacy threat into a security defense mechanism. By leveraging the unique characteristics of model inversion attacks, we can effectively detect backdoor attacks in federated learning systems. Our method exploits the fact that backdoored models exhibit distinct behavior patterns when subjected to model inversion techniques, allowing us to identify compromised participants in the federated learning process without requiring access to training data or model parameters. The Figure shows an example of a specific backdoor attack on an FL framework trained for face identification and authentication. To detect backdoor attacks in FL framework without the existing limitations, we propose a model learning performance-oriented detection mechanism. Instead of analyzing the similarity or distribution of model weights, we focus on evaluating model functionality, adopting the MI attack algorithm to first reconstruct data samples and then analyze model functionality based on the output.

Hierarchical Federated Learning with Group Local Differential Privacy

Hierarchical Federated Learning with SRS Mechanism

Federated learning (FL) is a distributed learning framework that allows clients to jointly train a model by uploading parameter updates rather than sharing local data. FL deployed on a client-edge-cloud hierarchical architecture, named Hierarchical Federated Learning (HFL), can accelerate model training and accommodate more clients with reduced communication cost via edge aggregation. Unfortunately, HFL suffers from privacy risks since the submitted parameters from clients are vulnerable to privacy attacks. To address this issue, we propose a novel Differential Privacy (DP) definition tailored for HFL, i.e., Group Local Differential Privacy (GLDP). We design the Sampling-Randomizing-Shuffling (SRS) mechanism to implement GLDP in HFL, where the sampling process is employed to achieve a stronger level of privacy protection with less noise added. By combining the randomized response and the shuffling mechanism, our proposed SRS mechanism can achieve client-level personalization within ρk-GLDP for privacy preservation while balancing model performance and privacy protection in HFL.

Contact

Lab Director: Dr. Honglu Jiang

Location: 105 Tallawanda Rd.
Department of Computer Science and Software Engineering
Miami University, Oxford, OH 45056

Email: jiangh34@miamioh.edu

Phone: 513-529-0353