Research Statement

Rajesh Kumar, PhD
Assistant Professor | Bucknell University, PA, USA

My research studies computational models of human behavioral signals and their applications in cybersecurity and natural language processing. The broader goal is to build trustworthy, human-centered intelligent systems grounded in measurable patterns of human interaction.

My research trajectory began in India with sensor localization in wireless sensor networks. I developed a novel algorithm for estimating the positions of deployed sensors using anchors with known locations, motivated by applications in GPS navigation and defense. During my doctoral studies, I shifted toward continuous identity verification based on behavioral signals from active and passive interactions with computing devices. This work examined walking, typing, and touch gestures as persistent biometric signals. It also introduced threat models, including mimicry, dictionary, and robotic attacks, enabling more realistic evaluation of continuous verification systems and challenging traditional assumptions. During my visiting and tenure-track appointments, I extended this research to automated signature verification and developed countermeasures against previously proposed adversaries, improving the robustness of behavioral authentication systems.

In parallel, I expanded behavioral biometrics research beyond identity and security applications. We developed a touchless typing system based on facial movement patterns that allows users to type by looking at a virtual keypad displayed on a monitor. This work created new interaction pathways for users with motor impairments and received Second Prize at the ACM Student Research Competition at the Richard Tapia Conference
(https://src.acm.org/winners/2024).

We later incorporated synthesized reading patterns into natural language processing models, improving performance on sentiment and sarcasm detection tasks (EACL 2023). More recently, we introduced a method for detecting plagiarism using keystroke dynamics, showing that typing behavior differs across bona fide writing, paraphrasing, and transcription modes (IJCB 2024, IJCB 2025). This work received the Best Paper Award at IEEE IJCB 2024 (https://ijcb2024.ieee-biometrics.org/award-winners/) and advances authorship attribution beyond traditional content-based detection.

With these developments, my research has consolidated into the following dimensions.

Behavioral Biometrics in Identity

This dimension develops continuous and context-aware authentication systems that secure digital platforms using human behavioral signals. Rather than treating authentication as a single login event, identity is modeled as an ongoing inference problem in which typing behavior, touch gestures, and motion signals provide continuous evidence of user presence. Early contributions introduced frameworks for incorporating contextual information into continuous verification and demonstrated the feasibility of behavioral authentication beyond static credentials.

Building on this foundation, my research examined adversarial robustness in behavioral authentication by systematically modeling attacks against gait and swiping biometrics. We developed treadmill-assisted mimicry and dictionary attacks, demonstrating how behavioral systems can be deceived under realistic adversarial settings. Subsequent work improved robustness through multimodal and context-aware learning. Later contributions used generative modeling and modality fusion to strengthen resilience against increasingly sophisticated attacks.

Integrity in the Age of Generative AI

This research addresses challenges of authenticity in digital communication and education arising from the rapid adoption of generative AI systems. As large language models blur distinctions between human- and machine-produced content, methods based solely on textual analysis are becoming less reliable. Our work instead evaluates how digital artifacts are produced rather than focusing only on what they contain.

We showed that typing behavior can distinguish genuine authorship from AI-assisted writing, earning the Best Paper Award at IEEE IJCB 2024. An extended study examined three writing modes: bona fide composition, transcription, and paraphrasing, providing a broader behavioral characterization of AI-assisted writing (incoming IEEE TBIOM 2026). This work was expanded to cheating detection in Korean, demonstrating cross-linguistic robustness. The resulting dataset also enabled analysis linking behavioral authorship signals to continuous authentication across varying levels of language model assistance and cognitive contexts. Ongoing work includes AI-assisted coding detection and deception-aware threat models.

Beyond educational settings, we applied keystroke dynamics to detecting deceptive accounts in online social networks without relying on intrusive content inspection. Complementary work examined structural and stylistic alignment between human and AI-generated text, analyzing how effectively large language models imitate individual writing styles. Together, these efforts position behavioral forensics as a complementary approach to authorship verification and online authenticity.

Human-Centered and Accessible AI

Traditional NLP systems rely primarily on textual features for tasks such as sentiment and sarcasm detection. In this work, we incorporated human behavioral signals as auxiliary information to improve model understanding. We introduced gaze patterns as an additional learning signal by injecting synthetically generated scanpaths that approximate reading behavior, enhancing model performance (EACL 2023). Extending this idea to interaction, we developed a touchless text-entry system that uses facial movements captured via Apple’s augmented reality framework. The system translates facial motion into written language, expanding access to computing for users with motor impairments (Tapia 2023).

Together, these efforts demonstrate how behavioral biometrics contribute not only to security and authorship verification but also to AI systems that better account for human attention, interaction, and accessibility.

Undergraduate Research Mentorship

A defining feature of my research program is the integration of undergraduate students as active contributors. I established the Behavioral Biometrics Research and Development Group (BRAG) as the primary environment for collaborative research and mentorship:
https://sites.google.com/view/kumar7/research/research-lab

The experimental nature of this work allows students to engage in research as early as their sophomore or junior years.

Over the past six years, more than thirty students have secured competitive summer funding through projects developed in my group. I work closely with students to design proposals and guide them through the full research lifecycle, including literature review, problem formulation, experimental design, data collection, machine learning implementation, and dissemination.

Projects in authentication, AI integrity, and human-centered AI naturally decompose into interconnected components, allowing students to contribute meaningfully while developing independence. Course projects frequently evolve into publishable research (ICMLA 2025). BRAG members have co-authored publications, received competitive funding and national recognition, including ACM SRC awards and CRA Outstanding Undergraduate Researcher Honorable Mention, and have gone on to graduate study or research careers. Mentorship is, therefore, a central mechanism through which innovation occurs in my program.

Taken together, my research uses human behavioral signals to improve security, authorship verification, and human–AI interaction. Across these areas, behavior serves as measurable evidence of identity and intent rather than relying solely on static credentials or textual content. Moving forward, I will continue developing behavioral methods that make intelligent systems more reliable, usable, and accessible while supporting sustained undergraduate participation in research.