AI7042 AI Reverse Engineering

Course Description

Time & Location

• Monday 15:00-17:45 @ College of Electronics Information and Applied Science 211-2

Objectives

• Understand how to analyze existing AI models in various aspects

• Gain insights into the design, functionality, and potential vulnerabilities of AI models

• Have a hands-on experience with AI reverse engineering

Structure

The course will consists of three parts:

• • Lecture: The lecturer will introduce the basic concepts of AI reverse engineering.

  • Seminar: Each student will present a review of recent research papers on AI reverse engineering.

  • Project: Each student will work on a small project regarding AI reverse engineering

Schedule

• Week  1 2025.9.01 Introduction

• Week  2 2025.9.08 Model Architecture Analysis

• Week  3 2025.9.15 Interpretability & Feature Analysis

• Week  4 2025.9.22 Paper Reviews

• Week  5 2025.9.29 Paper Reviews

• Week  6 2025.10.06 No Lecture

• Week  7 2025.10.13 Paper Reviews

• Week  8 2025.10.20 Paper Reviews (Remote)

• Week  9 2025.10.27 Model Extraction Attack

• Week 10 2025.11.03 Data Inference Attack

• Week 11 2025.11.10 Paper Reviews

• Week 12 2025.11.17 Paper Reviews

• Week 13 2025.11.24 Paper Reviews

• Week 14 2025.12.01 Paper Reviews

• Week 15 2025.12.08 Project Presentation (Final)

• Week 16 2025.12.15 Project Presentation (Final)

Presentation

• Part I: Analysis & Optimization

  • Week  4 2025.9.22

    • RepViT: Revisiting Mobile CNN From ViT Perspective, CVPR 2024, Muhammad Talha

    • DeepVID: Deep Visual Interpretation and Diagnosis for Image Classifiers via Knowledge Distillation, TVCG 2019, Pham Thanh Trung

    • AGAIN: Adversarial Training with Attribution Span Enlargement and Hybrid Feature Fusion, CVPR 2023, Tufail Hafiz Zahid

    • B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable, NeurIPS 2024, 김명철

    • Reverse Engineering Learned Optimizers: Mechanisms and Interpretations, NeurIPS 2021, 강은애

    • Towards Automated Circuit Discovery for Mechanistic Interpretability, NeurIPS 2023, 이슬찬

  • Week  5 2025.9.29

    • A ConvNet for the 2020s, CVPR 2022, 이태화

    • Scalable Image Coding for Humans and Machines, IEEE TIP 2022, 임달홍

    • This Looks Like That: Deep Learning for Interpretable Image Recognition ,NeurIPS2019, 옥윤승

    • Reverse-Engineering the Retrieval Process in GenIR Models, SIGIR 2025, 임준원

    • What Matters in Transformers? Not All Attention is Needed, arXiv 2024, 곽교린

    • Smoothquant: Accurate and efficient post-training quantization for large language models, ICML 2023, Faizan Rao

  • Week  7 2025.10.13

    • Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks, CVPR 2020, 최윤정

    • KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization, NeurIPS 2024, 한지윤

    • VanillaNet: the Power of Minimalism in Deep Learning, NeurIPS 2025, 조수현

    • Vision Transformers Need Registers, ICLR 2024, 진일성

    • LoRETTA: Low-Rank Economic Tensor-Train Adaptation for Ultra-Low-Parameter Fine-Tuning of Large Language Models, NAACL 2024, 김민환

    • On the Faithfulness of Vision Transformer Explanations, CVPR 2024, 이소연

  • Week  8 2025.10.20

    • Not All Adapters Matter: Selective Adapter Freezing for Memory-Efficient Fine-Tuning of Language Models, NAACL 2025, 박은주

    • Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas, ICML 2025, 김하늘

    • Quantized Feature Distillation for Network Quantization, AAAI 2023, 이정현

    • WWW: A Unified Framework for Explaining What, Where and Why of Neural Networks by Interpretation of Neuron Concepts. CVPR 2024, 안용현

    • CRAFT: Concept Recursive Activation FacTorization for Explainability, CVPR 2023, 조은기

    • What Does BERT Look At? An Analysis of BERT’s Attention, ArXiv 2019, 장하록

    • Subspace optimization for large language models with convergence guarantees, ICML 2025, Bold Chinguun

• Part II: Security & Threats

  • Week 11 2025.11.10

    • Locating and Editing Factual Associations in GPT, NeurIPS 2022,김원진

    • A Method to Facilitate Membership Inference Attacks in Deep Learning Models, NDSS 2025, 최용진

    • Gradient Inversion Attacks on Parameter-Efficient Fine-Tuning, CVPR 2025, 김정현

    • Transferable Structural Sparse Adversarial Attack Via Exact Group Sparsity Training, CVPR 2024, 김영웅

    • LoRA-Leak: Membership Inference Attacks Against LoRA Fine-tuned Language Models, arXiv 2025, 손소현

    • Implicit Bias Injection Attacks against Text-to-Image Diffusion Models, CVPR 2025, 나웅재

  • Week 12 2025.11.17

    • Membership Inference Attacks against Large Vision-Language Models, NeurIPS 2024, 이민재

    • Assessing Prompt Injection Risks in 200+ Custom GPTs, ICLR 2024 Workshop, 윤희균

    • Text Embeddings Reveal (Almost) As Much As Text, EMNLP 2023, 최용빈

    • AUTODAN: GENERATING STEALTHY JAILBREAK PROMPTS ON ALIGNED LARGE LANGUAGE MODELS, ICLR 2024, Jin Zhengxun

    • In-Context Unlearning: Language Models as Few Shot Unlearners, ICML 2024 (Poster), 강명구

    • Identifying Pre-training Data in LLMs: A Neuron Activation-Based Detection Framework, EMNLP 2025, 박재윤

  • Week 13 2025.11.24

  • Week 14 2025.12.01

Evaluation

The grade will be given according to the following grading percentages.

• Presentation 60%

• Project 30%

• Attendance 10%

Teaching Assistant