CS 59000BB: Situation Awareness, Adversarial ML and Explainable AI, Fall 2020

Home Reading List Slides Assignments Project

Time: Tue., Thu. 4:30pm - 5:45pm
Room: TBA
Instructor: Prof. Bharat Bhargava . Office: LWSN 2116F, Tel: 494-6013, Email: bbshail AT purdue.edu, Office hours: TBA and by appointment.
TA: TBA
Midterm Exam: TBA
Final Exam: TBA

Objective

Develop tools and systems that apply ML to real applications on real data and try to deal with attacks and explain the decisions of AI powered decision making. In addition, the seminar will discuss methods to make these tools more efficient and more accessible.

Teaching Material

Multi-modal data fusion, knowledge graphs, modeling context and situation awareness, user profiling and matching interests with streaming data ( sensors, text, tweets, video, news articles, emails, phone calls), pattern recognition, data mining, intelligent query processing. Machine learning models to connect user's need with data based on situation and context awareness.

Privacy-enabling frameworks for situational aware systems, Representing networks and knowledge graphs using graph databases, Graph analytics for enhancing search.

Machine learning in Data Cleaning, Video data understanding and data mining and identifying objects and events. Modeling user queries to build knowledge graphs to target data to users. AutoML and self-supervised learning: Deep learning to accelerate labeling of data and carefully tune hyper-parameters. Learn automatically and learn on unlabeled data to make machine learning accurate and efficient. Labeling and learning with incomplete or limited data.

Joint Modeling (Language Modeling, Multi-modal modeling, User Modeling), Open Information Extraction, Attribute Extraction, Relation Extraction, Similarity Learning, Graph Similarity, Pedestrian Attribute and Action Recognition, Intent Classification. Using tweets as the information shared by the people to visualize topic modeling, study subjectivity and to model the human emotions during the COVID-19 pandemic. Sharing information (e.g. personal opinions, some facts, news, status, etc.) on social media platforms which can be helpful to understand the various public behavior such as emotions, sentiments, and mobility during the ongoing pandemic and police encounters. Deep Learning approach for tweet classification for disaster management. Hierarchical deep learning model using text embedding via Crisis and GloVe, bidirectional LSTM (BLSTM), attention and convolution layers.

Attacks on ML models, training data, and streaming data. Preserving privacy and security by monitoring attacks, mitigating attacks, dealing with changing behavior of adversary, collaborative attacks, predicting attacks and intent. Deep learning based Programming Language Processing for Detecting Evasive Cyber-attacks, Trusted Classification under Poisoning attacks using Semantic Factors. Bias and Fairness in ML, what fairness means, causes that introduce unfairness in ML, Data-driven methods that unintentionally encode existing human biases and introduce new ones.

Defensive methods to protect Deep Neural Networks (DNNs). The success DNNs in image-related applications is threatened by their vulnerability to adversarial settings, including trojan and adversarial sample attacks. Detailed explanation of these types attacks and countering methods are covered.

Insider threats are considered one of the most serious and difficult problems to solve, given the privileges and information available to insiders to launch different types of attacks. Current security systems can record and analyze sequences from a deluge of log data, potentially becoming a tool to detect insider threats. The issue is that insiders mix the sequence of attack steps with valid actions, reducing the capacity of security systems to discern long sequences and programmatically detect the executed attacks. M Deep learning-based methods are introduced, which overcome the existing limitations and protect against these types of attacks.

Monitoring and learning of attacks on autonomous systems and cyber attribution. Enhancing Cognitive Autonomy through Deep learning. Autonomous systems, explanation of ML and actions of autonomous activity with human in the loop.

Attacks on space system protocols that use moving target defense and utilizing ML to identify intent and devising methods to mitigate.

Graph machine learning, graph structure to understand the social network and extract useful information to analyze social relationships. Build recommend systems. Deep learning in trading: Deep learning to track the market and make trading decisions.

Privacy preserving predictive modeling in edge networks, privacy issues in ML used in contact tracing.

Additional sources of knowledge

Outside speakers from MIT, CMU, Northrop Grumman, Sandia, Missouri Institute of Science and Technology, EPFL, University of Bern and Purdue (ECE, IE, CS) have been invited. Research papers will be assigned for reading and discussions in class. Dawn project at Stanford https://dawn.cs.stanford.edu/

Data sources:

Course Outline:

  1. Students will read/present papers and get involved in ongoing projects in Purdue.
  2. Students will learn about research in industry and projects funded at Purdue and develop research papers on a topic of interest based on their projects.

Assignments and Grading Policy:

The following assignments/exams etc. are planned.

Examination
  15%  
Research paper/report   10%
Five or Six assignments   25%
Class participation and discussion   10%
Class presentations   10%
Projects   30%

Projects

Class contribution

Some things that contribute to class participation:
 

Academic Dishonesty

One can not use any part of another person's work in his/her assignments and projects. If an overlap in any submission for grading is detected, an automatic grade of F in course will be assigned to the student and it will be reported to graduate school and deans. You are welcomed to discuss and learn from others. If you use any material in your homework, you have to put a reference to it. If you do a group project, specify individual contributions clearly at the time of submission and let me and TA know in advance of collaborating. Read the following links: Course Policies,   Academic Integrity.  
 

Comprehensive Reading list

Machine Learning

Adversarial Machine Learning

Situational Knowledge and Knowledge Graphs

Entity Resolution

Entity Resolution Explanations

Video Processing

Data Management For Video Streams

Attacks and Privacy in Distributed Systems

Talks

Covid-19 Conference

Other resources