Covert Task Embedding: Turning a DNN Into an Insider Agent Leaking Out Private Information

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Deep neural networks (DNNs) are increasingly used in sensitive applications.
• Existing security threats often focus on direct model manipulation or data breaches.
• A novel threat vector involves embedding hidden functionalities within DNNs.

Purpose of the Study:

• To introduce and define the Covert Task Embedding (CTE) attack.
• To demonstrate the feasibility of stealthily embedding and extracting sensitive information from DNNs.
• To highlight a new class of privacy and security risks in AI systems.

Main Methods:

• Developing a CTE attack methodology for DNNs.
• Training a face-based age estimation DNN to perform auxiliary gender and ethnicity classification tasks.
• Implementing a secret key mechanism for secure extraction of embedded task results.
• Evaluating the impact of the CTE attack on the primary task's accuracy.

Main Results:

• Successfully demonstrated the CTE attack in multiple experimental settings.
• Showed that private information (gender, ethnicity) can be reliably extracted.
• Confirmed that the auxiliary task extraction does not degrade the primary age estimation accuracy.
• Validated the stealthy nature of the CTE attack.

Conclusions:

• The CTE attack is a feasible and generalizable threat to DNNs.
• Current DNNs may be vulnerable to embedded malicious tasks.
• There is an urgent need for developing defenses against CTE attacks to protect sensitive data.