Building Trustworthy AI

Abstract: Auditability and regulatory compliance are increasingly required for deploying large language models (LLMs). Prior work typically targets isolated stages such as training or unlearning and lacks a unified mechanism for verifiable accountability across model updates. This paper presents AuditableLLM, a lightweight framework that decouples update execution from an audit-and-verification layer and records each update as a hash-chain-backed, tamper-evident audit trail. The framework supports parameter-efficient fine-tuning such as Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA), full-parameter optimization, continual learning, and data unlearning, enabling third-party verification without access to model internals or raw logs. Experiments on LLaMA-family models with LoRA adapters and the MovieLens dataset show negligible utility degradation (below 0.2% in accuracy and macro-F1) with modest overhead (3.4 ms/step; 5.7% slowdown) and sub-second audit validation in the evaluated setting. Under a simple loss-based membership inference attack on the forget set, the audit layer does not increase membership leakage relative to the underlying unlearning algorithm. Overall, the results indicate that hash-chain-backed audit logging can be integrated into practical LLM adaptation, update, and unlearning workflows with low overhead and verifiable integrity.

Abstract: Machine unlearning is becoming a practical requirement for deploying large language models (LLMs) under privacy, safety, and compliance constraints. In parallel, modern adaptation pipelines increasingly rely on parameter-efficient fine-tuning (PEFT), where a frozen backbone is paired with lightweight trainable modules such as Low-Rank Adaptation (LoRA). This survey reviews LoRA-based machine unlearning for LLMs, focusing on adapter-centric settings in which forgetting is achieved by operating primarily in PEFT module space rather than updating the full model. We formalize the LoRA unlearning problem under a frozen base model and present a taxonomy of existing methods, including gradient-based objectives, influence- and Fisher-informed updates, structural and continual-unlearning schemes, bounded-dynamics approaches, and pruning-enhanced variants. To relate these families, we introduce a unified adapter-space view that expresses seemingly disparate techniques as instances of a common update operator over adapter parameters. We further summarize evaluation protocols, metrics, and practical assumptions used in prior work, and provide a worked Task Of Fictitious Unlearning (TOFU) case study to illustrate utility–forgetting trade-offs across representative categories. Finally, we highlight open challenges such as repeated deletion requests, constrained data-access settings, federated or split-adapter deployments, and integration with auditing and compliance workflows.

Abstract: Large language models (LLMs) trained on decentralized data create a need for federated machine unlearning, where the influence of specific clients or authors is removed without retraining from scratch. Existing benchmarks such as TOFU target centralized author-level unlearning and do not address federated optimization or parameter-efficient adaptation. We introduce FedVILA, a LoRA-based federated unlearning framework that combines TOFU-style evaluation with importance-guided parameter updates. In our setup, a small subset of TOFU authors is assigned to a forget client, while the remaining authors are distributed over retain clients. Starting from a Llama 3.2 1B model trained via FedAvg with LoRA on all clients, FedVILA runs a federated unlearning phase where the forget client optimizes an Inverted Hinge Loss and retain clients continue language modeling. Using gradient-based importance to define a low-rank LoRA subspace, we report TOFU Truth Ratio and KS metrics, showing improved federated forgetting with largely preserved utility.

Project Overview: This project explores the intersection of computer graphics and computer vision in the browser. It renders a particle-based Christmas tree using Three.js, where each particle interacts with a physics system.

Key Features:

• Gesture Control: Integrated MediaPipe Hand Tracking to allow users to rotate and interact with the 3D scene using hand movements in real-time.
• Dynamic Particles: Custom shader materials and particle physics for visual effects.
• Photo Wall: An interactive floating photo gallery embedded within the 3D space.

Project Overview: Designing a resume that passes ATS (Applicant Tracking Systems) while maintaining high aesthetic standards can be challenging. This project provides a highly customizable XeLaTeX template focused on readability and structure.

Key Features:

• Custom Class: A dedicated resume.cls file that encapsulates complex formatting logic, keeping the main .tex file clean.
• Helper Macros: Pre-defined commands for education entries, skill lists, and contact information to ensure consistent vertical rhythm.
• Modern Aesthetics: Uses FontAwesome icons and a flat design language suitable for academic and engineering roles.