DSPLab advances a provenance-centric paradigm for AI trustworthiness, with focus on security, privacy, safety, explainability, and ethical norms for real-world deployment.

As AI systems increasingly influence high-stakes domains such as cybersecurity, finance, healthcare, and autonomous systems, the core challenge is no longer only improving accuracy but ensuring that AI systems remain robust, safe, trustworthy, and accountable throughout their lifecycle. The key lies in provenance: understanding and tracing the origins, lineage of transformations, and influence pathways that shape a model’s behavior. Traditional AI evaluation relies largely on black-box testing, observing outputs without visibility into the internal processes that produced them. This leaves critical blind spots against threats such as data poisoning, backdoor attacks, adversarial manipulation, and unsafe or unintended model behavior. DSPLab introduces fine-grained observability into the AI pipeline by tracking the lifecycle history of data, training dynamics, parameter updates, and inference-time information flows. Through this lens, we develop theoretical foundations, algorithms, and systems that make robustness and safety measurable, explainable, and auditable, enabling attack detection, forensic analysis, accountability, and automated model repair. Our vision is to establish provenance as a foundational layer for AI security and AI safety, transforming AI from opaque systems into observable and auditable infrastructures where model decisions can be traced, inspected, and verified for responsible deployment.