Secure Machine Learning Training Overview
Attacks on machine learning applications are gaining momentum and protecting against a machine learning (ML) breach is essential. This Machine Learning Security training course teaches developers the skills they need to protect their ML applications. Students learn specialized secure coding skills and how to avoid the security pitfalls of the Python programming language.
Note: To ensure ample one-on-one engagement with the instructor, this class is capped at 12 people, overriding Accelebrate’s default cap of 15.
Location and Pricing
Accelebrate offers instructor-led enterprise training for groups of 3 or more online or at your site. Most Accelebrate classes can be flexibly scheduled for your group, including delivery in half-day segments across a week or set of weeks. To receive a customized proposal and price quote for private corporate training on-site or online, please contact us.
In addition, some courses are available as live, instructor-led training from one of our partners.
Objectives
- Understand essential cyber security concepts
- Learn about various aspects of machine learning security
- Discover the possible attacks and defense techniques in adversarial machine learning
- Identify vulnerabilities and their consequences
- Learn the security best practices in Python
- Understand Input validation approaches and principles
- Manage vulnerabilities in third-party components
- Understand how cryptography can support application security
- Learn how to use cryptographic APIs correctly in Python
- Understand security testing methodology and approaches
- Be familiar with common security testing techniques and tools
Prerequisites
Students should be Python developers working on machine learning systems.
Outline
Expand All | Collapse All
Cyber Security Basics
- What is security?
- Threat and risk
- Cyber security threat types
- Consequences of insecure software
- Constraints and the market
- The dark side
- Categorization of bugs
- The Seven Pernicious Kingdoms
- Common Weakness Enumeration (CWE)
- CWE Top 25 Most Dangerous Software Errors
- Vulnerabilities in the environment and dependencies
Cyber Security in Machine Learning
- ML-specific cyber security considerations
- What makes machine learning a valuable target?
- Possible consequences
- Inadvertent AI failures
- Some real-world abuse examples
- ML threat model
- Creating a threat model for machine learning
- Machine learning assets
- Security requirements
- Attack surface
- Attacker model – resources, capabilities, goals
- Confidentiality threats
- Integrity threats (model)
- Integrity threats (data, software)
- Availability threats
- Dealing with AI/ML threats in software security
Using ML in Cyber Security
- Static code analysis and ML
- ML in fuzz testing
- ML in anomaly detection and network security
- Limitations of ML in security
Malicious Use of AI and ML
- Social engineering attacks and media manipulation
- Vulnerability exploitation
- Malware automation
- Endpoint security evasion
Adversarial Machine Learning
- Threats against machine learning
- Attacks against machine learning integrity
- Poisoning attacks
- Poisoning attacks against supervised learning
- Poisoning attacks against unsupervised and reinforcement learning
- Evasion attacks
- Common white-box evasion attack algorithms
- Common black-box evasion attack algorithm
- Transferability of poisoning and evasion attacks
- Some defense techniques against adversarial samples
- Adversarial training
- Defensive distillation
- Gradient masking
- Feature squeezing
- Using reformers on adversarial data
- Caveats about the efficacy of current adversarial defenses
- Simple practical defenses
- Attacks against machine learning confidentiality
- Model extraction attacks
- Defending against model extraction attacks
- Model inversion attacks
- Defending against model inversion attacks
Denial of Service
- Denial of Service
- Resource exhaustion
- Cash overflow
- Flooding
- Algorithm complexity issues
- Denial of service in ML
- Accuracy reduction attacks
- Denial-of-information attacks
- Catastrophic forgetting in neural networks
- Resource exhaustion attacks against ML
- Best practices for protecting availability in ML systems
Input Validation Principles
- Blacklists and whitelists
- Data validation techniques
- What to validate – the attack surface
- Where to validate – defense in depth
- How to validate – validation vs transformations
- Output sanitization
- Encoding challenges
- Validation with regex
- Regular expression denial of service (ReDoS)
- Dealing with ReDoS
Injection
- Injection principles
- Injection attacks
- SQL injection
- SQL injection basics
- Attack techniques
- Content-based blind SQL injection
- Time-based blind SQL injection
- SQL injection best practices
- Input validation
- Parameterized queries
- Additional considerations
- SQL injection and ORM
- Code injection
- Code injection via input()
- OS command injection
- General protection best practices
Integer Handling Problems
- Representing signed numbers
- Integer visualization
- Integers in Python
- Integer overflow
- Integer overflow with ctypes and NumPy
- Other numeric problems
Files and Streams
- Path traversal
- Path traversal-related examples
- Additional challenges in Windows
- Virtual resources
- Path traversal best practices
- Format string issues
Unsafe Native Code
- Native code dependence
- Best practices for dealing with native code
Input Validation in Machine Learning
- Misleading the machine learning mechanism
- Sanitizing data against poisoning and RONI
- Code vulnerabilities causing evasion, misprediction, or misclustering
- Typical ML input formats and their security
Security Features
- Authentication
- Authentication basics
- Multi-factor authentication
- Authentication weaknesses - spoofing
- Password management
- Information exposure
- Exposure through extracted data and aggregation
- Privacy violation
- System information leakage
- Information exposure best practices
Time and State
- Race conditions
- File race condition
- Avoiding race conditions in Python
- Mutual exclusion and locking
- Synchronization and thread safety
Errors
- Error handling
- Returning a misleading status code
- Information exposure through error reporting
- Exception handling
- In the except, catch block. And now what?
- Empty catch block
- The danger of assert statements
Using Vulnerable Components
- Assessing the environment
- Hardening
- Malicious packages in Python
- Vulnerability management
- Patch management
- Vulnerability management
- Bug bounty programs
- Vulnerability databases
- Vulnerability rating – CVSS
- DevOps, the build process and CI / CD
- Dependency checking in Python
- ML Supply Chain Risks
- Common ML system architectures
- ML system architecture and the attack surface
- Protecting data in transit – transport layer security
- Protecting data in use – homomorphic encryption
- Protecting data in use – differential privacy
- Protecting data in use – multi-party computation
- ML frameworks and security
- General security concerns about ML platforms
- TensorFlow security issues and vulnerabilities
Cryptography for Developers
- Cryptography basics
- Cryptography in Python
- Elementary algorithms
- Random number generation
- Hashing
- Confidentiality protection
- Homomorphic encryption
- Basics of homomorphic encryption
- Types of homomorphic encryption
- FHE in machine learning
- Integrity protection
- Message Authentication Code (MAC)
- Digital signature
- Public Key Infrastructure (PKI)
- Some further key management challenges
- Certificates
Security Testing
- Security testing methodology
- Security testing – goals and methodologies
- Overview of security testing processes
- Threat modeling
- Security testing techniques and tools
- Code analysis
- Dynamic analysis
Wrap Up
- Secure coding principles
- Principles of robust programming by Matt Bishop
- Secure design principles of Saltzer and Schröder
- And now what?
- Software security sources and further reading
- Python resources
- Machine learning security resources
Training Materials
All attendees receive comprehensive courseware.
Software Requirements
Attendees will not need to install any software on their computers for this class. The class will be conducted in a remote environment that Accelebrate will provide; students will only need a local computer with a web browser and a stable Internet connection. Any recent version of Microsoft Edge, Mozilla Firefox, or Google Chrome will work well.
Machine Learning Security Webinar
In this ML Security Webinar, one of our senior secure code trainers discusses the ways your systems may be vulnerable to attacks and what you can do to avoid them.
