Skip to main content

Deep learning, represented by large language models, is revolutionizing human lives. However, trustworthiness threats in deep learning widely exist, posing great challenges to AI safety, security, and reliability. This course introduces state-of-the-art frontiers on deep learning research for a wide range of trustworthiness issues, including threat discovery, mitigation, and certification methods through seminar-style presentations and hands-on projects.

This is a seminar-style course for trustworthy deep learning. The first half of the course is an overview of deep learning and preliminaries for trustworthy AI methods, including training of neural networks, common neural network architectures, large language models, the definition of AI attacks, defences, and certification and verification in the context of AI. The second half of the course visits representative and recent research papers in the field through student presentations, covering topics like evasion attacks and defences, robustness certification, differential privacy, membership inference attacks, watermarks, detection of AI-generated contents, machine unlearning, prompt injection attacks, model stealing, and finetuning attacks.

Prerequisites #

There is no formal pre-requisite. Background in algorithms, calculus, linear algebra (e.g., MATH 151, MATH 152, MATH 232, CMPT 225), CMPT 410/726 Machine Learning strongly recommended. It is also recommended to have a background in CMPT 412/762 Computer Vision and CMPT 713 NLP.

Textbook and Reading Materials #

There is no primary reference material. We will read an assortment of research papers during lectures.

  • Deep Learning Book
    • By Ian Goodfellow, Yoshua Bengio, and Aaron Courville
    • Recommended for students to gain background in deep learning before taking the course.
  • Online course Intro to ML Safety
    • By Dan Hendrycks at the Center for AI Safety
    • Optional, advanced reading for interested students
    • A well-developed course recommended for those who want to learn general machine learning safety from a systematic and interdisciplinary perspective.

Grading #

10% Homework 0 (raw score) + 40% course project (1.1 × raw score with no cap) + 30% paper presentation + 20% notes of peer evaluation and summary

Schedule and Syllabus #

Syllabus

Slides will be updated as the term progresses. All slides are available in this OneDrive folder. The slides are password encrypted - password posted on Canvas.

Week Date Topics (Tentative) Assignment & Due Reading
Week 1 (1/5 - 1/11) Tue (1/7) 2h (Lecture) Syllabus, Introduction to Deep Learning I, Homework 0 Explained Homework 0 Release See References in slides
  Thur (1/9) 1h (Lecture) Introduction to Deep Learning II   See References in Slides
Week 2 (1/12 - 1/18) Tue (1/14) 2h (Lecture) Introduction to Deep Learning III   See References in Slides
  Thur (1/16) 1h (Lecture) Course Presentation Instructions; Trustworthy Deep Learning Overview Presentation Signing-up Sheet Release
Homework 0 Due (1/18)
See References in Slides
Week 3 (1/19 - 1/25) Tue (1/21) 2h (Lecture) Robustness Threats in Deep Learning - Attacks I   See References in Slides
  Thur (1/23) 1h (Lecture) Robustness Threats in Deep Learning - Attacks II   See References in Slides
Week 4 (1/26 - 2/1) Tue (1/28) 2h (Lecture) Robustness Threats in Deep Learning - Defenses   See References in Slides
  Thur (1/30) 1h (Lecture) Robustness Threats in Deep Learning - Certification I   See References in Slides
Week 5 (2/2 - 2/8) Tue (2/4) 2h Class cancelled    
  Thur (2/6) 1h (Lecture) Robustness Threats in Deep Learning - Certification I (Cont.d) & II Presentation Signing-up Due (2/8) See References in Slides
Week 6 (2/9 - 2/15) Tue (2/11) 2h (Lecture) Robustness Threats in Deep Learning - Certification II (Cont.d) Course Project Release See References in Slides
  Thur (2/13) 1h (Presentation) Privacy Attacks
Membership Inference Attacks Against Machine Learning Models (35 min presentation).
Notes Submission Required 1802.08232
1702.07464
Week 7 (2/16 - 2/22) Reading Break      
Week 8 (2/23 - 3/1) Tue (2/25) 2h (Presentation) Privacy Attacks & Differential Privacy
Deep Leakage from Gradients (15 min presentation).
Differentially Private Synthetic Data via Foundation Model APIs 1: Images & Differentially Private Synthetic Data via Foundation Model APIs 2: Text (35 min presentation)
Flocks of Stochastic Parrots: Differentially Private Prompt Learning for Large Language Models (15 min presentation).
Notes Submission Required 2301.13188
2404.17399
2112.03570
2202.07646
2302.00539
1607.00133
1702.07476
2106.02848
1610.05755
1805.10559
1802.08908
2110.06500
2110.05679
2212.06470
2205.10683
  Thur (2/27) 1h Class cancelled    
Week 9 (3/2 - 3/8) Tue (3/4) 2h (Presentation) Copyright, Unlearning, Model Stealing
Machine Unlearning (35 min presentation).
Protecting Artists from Style Mimicry by Text-to-Image Models (25 min presentation).
Extracting Training Data from Large Language Models (25 min presentation).
Notes Submission Required Towards Making Systems Forget with Machine Unlearning
2109.13398
2403.06634
2404.03233
1911.03030
2303.07345
  Thur (3/6) 1h (Presentation) Fairness
A Survey on Bias and Fairness in Machine Learning (35 min presentation).
Notes Submission Required 1906.08386
2205.15494
2002.10312
Week 10 (3/9 - 3/15) Tue (3/11) 2h (Presentation) Data Poisoning Attacks, Fairness
Fairness Without Demographics in Repeated Loss Minimization (15 min presentation).
(Lecture) Poisoning & Backdoor Attacks for Deep Neural Networks
Notes Submission Required 1804.00792
2302.10149
Trojaning Attack on Neural Networks
Robust Logistic Regression and Classification
1706.03691
Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
2006.14768
Exploring the Orthogonality and Linearity of Backdoor Attacks
  Thur (3/13) 1h (Presentation) Data Valuation
Towards Efficient Data Valuation Based on the Shapley Value (15 min presentation).
Studying Large Language Model Generalization with Influence Functions (15 min presentation).
Notes Submission Required Data Shapley: Equitable Valuation of Data for Machine Learning
1703.04730
Week 11 (3/16 - 3/22) Tue (3/18) 2h (Lecture) Large Language Models: Overview   CMU 11-711 ANLP
  Thur (3/20) 1h (Lecture) LLM Trustworthiness: Overview   DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models
AI Risk Management Should Incorporate Both Safety and Security
Recommendations for Technical AI Safety Research Directions
Week 12 (3/23 - 3/29) Tue (3/25) 2h (Presentation) LLM Alignment Tuning
Training Language Models to Follow Instructions with Human Feedback (25 min presentation).
Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback (25 min presentation).
Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations (25 min presentation).
SimPO: Simple Preference Optimization with a Reference-Free Reward (15 min presentation).
Notes Submission Required 2305.18290
Safety Alignment Should be Made More Than Just a Few Tokens Deep
2404.12358
2208.03274
  Thur (3/27) 1h (Presentation) LLM Safety Benchmarks
TruthfulQA: Measuring How Models Mimic Human Falsehoods (15 min presentation).
Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (15 min presentation).
Notes Submission Required 2402.04249
2308.03825
2406.14598
2111.02840
Week 13 (3/30 - 4/5) Tue (4/1) 2h (Presentation) LLM Prompting and Prompt Injection
Jailbroken: How Does LLM Safety Training Fail? (25 min presentation).
Universal and Transferable Adversarial Attacks on Aligned Language Models (15 min presentation).
Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks (15 min presentation).
Jailbreaking Black Box Large Language Models in Twenty Queries (15 min presentation).
Notes Submission Required 2306.15447
2401.06373
LLM-Fuzzer: Scaling Assessment of Large Language Model Jailbreaks
2310.06987
2306.13213
2404.01318
2405.09113
  Thur (4/3) 1h (Presentation) LLM Finetuning Attacks and Defenses
Locking Down the Finetuned LLMs Safety (15 min presentation).
Poisoning Attacks against Support Vector Machines (15 min presentation) - adjusted due to presenter absence.
Notes Submission Required
Course project Due (4/5)
2409.18169
2310.03693
2406.20053
2404.01099
2405.16833
Week 14 (4/6 - 4/12) Tue (4/8) 2h (Lecture) Course Project Discussion, Closing Remarks    
Week 15 (4/13-4/19) Fri (4/19) Grade Released    

Extended Topics #

Trustworthy deep learning is a broad area. Some important topics are not covered in lectures and presentations due to the limited time frame. Some of them are listed below.

  • LLM Hallucination
  • Risks of LLM agents
  • Reward hacking and goal misspecification in RL and RLHF
  • Social-economic Impact with Generative AI

Assignments and Project #

  • Homework 0
    • Deadline: 23:59, Jan 18, 2025
    • Score released (Jan 28, 2025)
  • Presentation:
  • Course Project:
    • Deadline: Apr 5, 2025
  • Note Submission:
    • Submission links dynamically released on Canvas
    • Only for student presentation dates
    • Due 7 days after each presentation date
    • Submit on Canvas
    • Up to 3 exemptions

Information Platform #

Ethics Statement #

This course will include topics related computer security and privacy. As part of this investigation we may cover technologies whose abuse could infringe on the rights of others. As computer scientists, we rely on the ethical use of these technologies. Unethical use includes circumvention of an existing security or privacy mechanisms for any purpose, or the dissemination, promotion, or exploitation of vulnerabilities of these services. Any activity outside the letter or spirit of these guidelines will be reported to the proper authorities and may result in dismissal from the class and possibly more severe academic and legal sanctions.

Academic Integrity Policy #

  • Some examples of unacceptable behaviour in homeworks and course projects:
    • Handing in assignments that are not 100% your own work (in design, implementation, wording, etc.), without proper citation. There must be a README file in your submission with citations to any external code used.
    • Sharing code fragments with others in class (for group project, with others who are not in the same group) is not allowed.
    • Keep discussions to high level information rather than specific code hints.
    • Copying and then obfuscating code is a serious academic honesty violation.
    • Submitting work that has been submitted before, for any course at any institution.
  • If you are unclear on what academic honesty is, see Simon Fraser University’s Policy S10-01.
  • All instances of academic dishonesty will be dealt with very severely.
  • In general, minimum requested penalties will be as follows:
    • For assignments and course projcets: a mark of -50% on the assignment. So, academic dishonesty on an assignment worth 10% of your final mark will result in a zero on the assignment, and a penalty of 5% from your final grade.
  • Please note that these are minimum penalties. At the instructor’s option, more severe penalties may be given/requested. All instances of academic dishonesty will be noted on your University record.
  • The instructor may use an automated service that will check for plagiarism.

Acknowledgement #

The course is developed from CS562 and CS598GS at UIUC. Part of the content is adapted from Intro to ML Safety. Some course policies are developed from CMPT 413 Natural Language Processing.