MiDA: Membership inference attacks against domain adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Domain adaptation is crucial for training neural networks with insufficient data.
• Existing domain adaptation techniques may inadvertently expose sensitive training data.
• Membership inference attacks pose a privacy risk by revealing if a data sample was used in training.

Purpose of the Study:

• To investigate the vulnerability of domain adaptation models to privacy breaches.
• To propose a novel membership inference attack tailored for domain adaptation.
• To assess the effectiveness of the attack in practical, black-box scenarios.

Main Methods:

• Developed a membership inference attack leveraging knowledge of an additional source domain.
• Exploited distributional similarities between target and source domains for attack.
• Evaluated the attack's performance on object and digit recognition tasks.
• Designed the attack for scenarios where model details are inaccessible (black-box).

Main Results:

• The proposed attack achieved high efficiency and accuracy in inferring membership information.
• Demonstrated successful attacks against domain adaptation models in object and digit recognition.
• The attack maintained a high success rate even without access to model parameters or architecture.

Conclusions:

• Domain adaptation models are susceptible to sophisticated membership inference attacks.
• The proposed attack effectively breaches training data privacy in domain adaptation settings.
• This research highlights the need for enhanced privacy-preserving techniques in domain adaptation.