LLM03: Training Data Poisoning

The starting point of any machine learning approach is training data, simply “raw text”. To be highly capable (e.g., have linguistic and world knowledge), this text should span a broad range of domains, genres and languages. A large language model uses deep neural networks to generate outputs based on patterns learned from training data.

Training data poisoning refers to manipulation of pre-training data or data involved within the fine-tuning or embedding processes to introduce vulnerabilities (which all have unique and sometimes shared attack vectors), backdoors or biases that could compromise the model’s security, effectiveness or ethical behavior. Poisoned information may be surfaced to users or create other risks like performance degradation, downstream software exploitation and reputational damage. Even if users distrust the problematic AI output, the risks remain, including impaired model capabilities and potential harm to brand reputation.

  • Pre-training data refers to the process of training a model based on a task or dataset.
  • Fine-tuning involves taking an existing model that has already been trained and adapting it to a narrower subject or a more focused goal by training it using a curated dataset. This dataset typically includes examples of inputs and corresponding desired outputs.
  • The embedding process is the process of converting categorical data (often text) into a numerical representation that can be used to train a language model. The embedding process involves representing words or phrases from the text data as vectors in a continuous vector space. The vectors are typically generated by feeding the text data into a neural network that has been trained on a large corpus of text.

Data poisoning is considered an integrity attack because tampering with the training data impacts the model’s ability to output correct predictions. Naturally, external data sources present higher risk as the model creators do not have control of the data or a high level of confidence that the content does not contain bias, falsified information or inappropriate content.

Common Examples of Vulnerability

  1. A malicious actor, or a competitor brand intentionally creates inaccurate or malicious documents which are targeted at a model’s pre-training, fine-tuning data or embeddings. Consider both Split-View Data Poisoning and Frontrunning Poisoning attack vectors for illustrations.
    • The victim model trains using falsified information which is reflected in outputs of generative AI prompts to it’s consumers.
  2. A malicious actor is able to perform direct injection of falsified, biased or harmful content into the training processes of a model which is returned in subsequent outputs.
  3. An unsuspecting user is indirectly injecting sensitive or proprietary data into the training processes of a model which is returned in subsequent outputs.
  4. A model is trained using data which has not been verified by its source, origin or content in any of the training stage examples which can lead to erroneous results if the data is tainted or incorrect.
  5. Unrestricted infrastructure access or inadequate sandboxing may allow a model to ingest unsafe training data resulting in biased or harmful outputs. This example is also present in any of the training stage examples.
    • In this scenario, a users input to the model may be reflected in the output to another user (leading to a breach), or the user of an LLM may receive outputs from the model which are inaccurate, irrelevant or harmful depending on the type of data ingested compared to the model use-case (usually reflected with a model card).

Whether a developer, client or general consumer of the LLM, it is important to understand the implications of how this vulnerability could reflect risks within your LLM application when interacting with a non-proprietary LLM to understand the legitimacy of model outputs based on it's training procedures. Similarly, developers of the LLM may be at risk to both direct and indirect attacks on internal or third-party data used for fine-tuning and embedding (most common) which as a result creates a risk for all it's consumers.

How to Prevent

  1. Verify the supply chain of the training data, especially when sourced externally as well as maintaining attestations via the "ML-BOM" (Machine Learning Bill of Materials) methodology as well as verifying model cards.
  2. Verify the correct legitimacy of targeted data sources and data contained obtained during both the pre-training, fine-tuning and embedding stages.
  3. Verify your use-case for the LLM and the application it will integrate to. Craft different models via separate training data or fine-tuning for different use-cases to create a more granular and accurate generative AI output as per it's defined use-case.
  4. Ensure sufficient sandboxing through network controls are present to prevent the model from scraping unintended data sources which could hinder the machine learning output.
  5. Use strict vetting or input filters for specific training data or categories of data sources to control volume of falsified data. Data sanitization, with techniques such as statistical outlier detection and anomaly detection methods to detect and remove adversarial data from potentially being fed into the fine-tuning process.
  6. Adversarial robustness techniques such as federated learning and constraints to minimize the effect of outliers or adversarial training to be vigorous against worst-case perturbations of the training data.
    • An "MLSecOps" approach could be to include adversarial robustness to the training lifecycle with the auto poisoning technique.
    • An example repository of this would be Autopoison testing, including both attacks such as Content Injection Attacks (“(attempting to promote a brand name in model responses”) and Refusal Attacks (“always making the model refuse to respond”) that can be accomplished with this approach.
  7. Testing and Detection, by measuring the loss during the training stage and analyzing trained models to detect signs of a poisoning attack by analyzing model behavior on specific test inputs.
    • Monitoring and alerting on number of skewed responses exceeding a threshold.
    • Use of a human loop to review responses and auditing.
    • Implement dedicated LLMs to benchmark against undesired consequences and train other LLMs using reinforcement learning techniques.
    • Perform LLM-based red team exercises or LLM vulnerability scanning in the testing phases of the LLM lifecycle.

Example Attack Scenarios

  1. The LLM generative AI prompt output can mislead users of the application which can lead to biased opinions, followings or even worse, hate crimes etc.
  2. If the training data is not correctly filtered and|or sanitized, a malicious user of the application may try to influence and inject toxic data into the model for it to adapt to the biased and false data.
  3. A malicious actor or competitor intentionally creates inaccurate or malicious documents which are targeted at a model’s training data in which is training the model at the same time based on inputs. The victim model trains using this falsified information which is reflected in outputs of generative AI prompts to it's consumers.
  4. The vulnerability Prompt Injection could be an attack vector to this vulnerability if insufficient sanitization and filtering is performed when clients of the LLM application input is used to train the model. I.E, if malicious or falsified data is input to the model from a client as part of a prompt injection technique, this could inherently be portrayed into the model data.