Security and Behavioral AI Projects

Our security-focused research explores behavioral biometrics, continuous authentication, and adversarial robustness to strengthen user privacy and trust in modern computing systems. By harnessing motion sensors, touch data, and deep learning, we create lightweight, real-time authentication frameworks that safeguard smartphones even under adversarial conditions. These efforts extend to dynamic ensemble learning and decision-level fusion, enabling adaptive and explainable AI models across high-risk domains like cybersecurity, finance, and mobile platforms.

Securing Interpretable Deep Learning Systems Adversarial Threats, Stealthy Attacks, and Defense Mechanisms

The growing integration of deep learning (DL) models into high-stakes domains, such as healthcare, finance, and autonomous systems, has made interpretability a cornerstone of trustworthy AI. Interpretable Deep Learning Systems (IDLSes), which combine powerful neural networks with interpretation models, aim to provide transparency into the decision-making process. However, the assumption that interpretation inherently adds security has recently been challenged.

Comprehensive Evaluation of Adversarial Robustness in Deep Learning Architecture, Diversity, and Defense Analysis

Adversarial attacks pose a serious challenge to the reliability and security of deep learning (DL) models. These attacks, often crafted by introducing imperceptible perturbations to input data, can cause models to make incorrect predictions with high confidence. As a result, understanding and mitigating such threats has become a critical area of research in the field of trustworthy AI. Defenses against adversarial attacks range from input preprocessing and adversarial training to robust model design, yet no single approach has proven universally effective.

Robust Malware Detection in Adversarial Environments Analysis, Evaluation, and Defense Strategies

The dynamic evolution of malware, combined with increasingly sophisticated evasion techniques such as packing, obfuscation, and polymorphism, presents a significant challenge to conventional security mechanisms. Machine learning (ML)-based malware detection systems are widely adopted for their ability to generalize and automate malware identification, yet they remain susceptible to adversarial threats. InfoLab at SKKU investigates robust, interpretable detection pipelines—spanning spectral control-flow-graph analysis, the effects of packing on ML detectors, and visualization-based feature fusion—to identify evasive and morphed malware across desktop and mobile platforms.

Behavioral Biometrics for Continuous and Adversarially Robust User Authentication on Smartphones

Traditional authentication methods—such as passwords, PINs, and even biometric systems (fingerprint, facial recognition)—typically secure mobile devices only at the point of entry. However, they fail to offer protection throughout a session, leaving devices vulnerable to unauthorized access when unattended. To bridge this security gap, the research group InfoLab at Sungkyunkwan University (SKKU) has led a series of studies on continuous, sensor-based, and adversarially-aware user authentication mechanisms.

Robust Federated Learning: Defending Distributed AI Against Poisoning Attacks Poisoning Attacks Under Client Data Heterogeneity

Federated learning lets many clients collaboratively train a shared model without exchanging raw data, making it ideal for privacy-sensitive domains like healthcare and finance. But because the server never sees client data, malicious participants can quietly corrupt the global model through data- and model-poisoning attacks. InfoLab at SKKU investigates how robust federated learning really is once clean, identically distributed data is no longer assumed—studying how poisoning attacks behave under realistic client data heterogeneity (non-IID settings) and how robust-aggregation defenses must adapt when honest clients already look very different from one another.