SEEK: A simple defense to model hijacking attack.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Cybersecurity

Background:

• Model hijacking attacks pose a significant training-time threat to NLP systems.
• These attacks enable dual-functionality, allowing malicious tasks alongside original objectives.
• Existing defenses are insufficient against these sophisticated threats.

Purpose of the Study:

• To propose SEEK, a novel algorithm for defending against model hijacking.
• To neutralize attacks by disrupting the attacker's inference procedure.
• To maintain the functional integrity of NLP models post-defense.

Main Methods:

• SEEK employs a proactive defense mechanism to identify and eliminate attacker-embedded prompt words.
• A fill-mask model is utilized to substitute deleted terms, preventing semantic degradation.
• The algorithm combines precise deletion with contextual synonym substitution for robust security.

Main Results:

• SEEK reduced the model hijacking attack success rate by up to 75%.
• Original task performance was maintained across diverse model architectures.
• Demonstrated a threefold improvement in defense efficacy compared to ONION and RAP.

Conclusions:

• SEEK offers a simple yet effective defense against model hijacking attacks.
• The proposed method successfully neutralizes attacks while preserving model utility.
• SEEK represents a significant advancement in securing NLP systems against training-time threats.