CWE-1039

Inadequate Detection or Handling of Adversarial Input Perturbations in Automated Recognition Mechanism
AI Translation Available

The product uses an automated mechanism such as machine learning to recognize complex data inputs (e.g. image or audio) as a particular concept or category, but it does not properly detect or handle inputs that have been modified or constructed in a way that causes the mechanism to detect a different, incorrect concept.

Status
incomplete
Abstraction
class
AI/ML

When techniques such as machine learning are used to automatically classify input streams, and those classifications are used for security-critical decisions, then any mistake in classification can introduce a vulnerability that allows attackers to cause the product to make the wrong security decision or disrupt service of the automated mechanism. If the mechanism is not developed or 'trained' with enough input data or has not adequately undergone test and evaluation, then attackers may be able to craft malicious inputs that intentionally trigger the incorrect classification.

Targeted technologies include, but are not necessarily limited to:

- automated speech recognition

- automated image recognition

- automated cyber defense

- Chatbot, LLMs, generative AI

For example, an attacker might modify road signs or road surface markings to trick autonomous vehicles into misreading the sign/marking and performing a dangerous action. Another example includes an attacker that crafts highly specific and complex prompts to 'jailbreak' a chatbot to bypass safety or privacy mechanisms, better known as prompt injection attacks.

Common Consequences

integrity availability confidentiality other
Impacts
bypass protection mechanism dos: resource consumption (other) dos: instability read application data varies by context

Detection Methods

dynamic analysis with manual results interpretation architecture or design review

Potential Mitigations

Phases:
architecture and design implementation integration
Descriptions:
• Algorithmic modifications such as model pruning or compression can help mitigate this weakness. Model pruning ensures that only weights that are most relevant to the task are used in the inference of incoming data and has shown resilience to adversarial perturbed data.
• Consider implementing model hardening to fortify the internal structure of the algorithm, including techniques such as regularization and optimization to desensitize algorithms to minor input perturbations and/or changes.
• Consider implementing adversarial training, a method that introduces adversarial examples into the training data to promote robustness of algorithm at inference time.
• Consider implementing multiple models or using model ensembling techniques to improve robustness of individual model weaknesses against adversarial input perturbations.
• Incorporate uncertainty estimations into the algorithm that trigger human intervention or secondary/fallback software when reached. This could be when inference predictions and confidence scores are abnormally high/low comparative to expected model performance.
• Reactive defenses such as input sanitization, defensive distillation, and input transformations can all be implemented before input data reaches the algorithm for inference.
• Consider reducing the output granularity of the inference/prediction such that attackers cannot gain additional information due to leakage in order to craft adversarially perturbed data.