SoK: Reducing the Vulnerability of Fine-tuned Language Models to Membership Inference Attacks

Natural language processing models have experienced a significant upsurge in recent years, with numerous applications being built upon them. Many of these applications require fine-tuning generic base models on customized, proprietary datasets. This fine-tuning data is especially likely to contain personal or sensitive information about individuals, resulting in increased privacy risk. Membership inference attacks are the most commonly employed attack to assess the privacy leakage of a machine learning model. However, limited research is available on the factors that affect the vulnerability of language models to this kind of attack, or on the applicability of different defense strategies in the language domain. We provide the first systematic review of the vulnerability of fine-tuned large language models to membership inference attacks, the various factors that come into play, and the effectiveness of different defense strategies. We find that some training methods provide significantly reduced privacy risk, with the combination of differential privacy and low-rank adaptors achieving the best privacy protection against these attacks.

Improved Membership Inference Attacks Against Language Classification Models

Artificial intelligence systems are prevalent in everyday life, with use cases in retail, manufacturing, health, and many other fields. With the rise in AI adoption, associated risks have been identified, including privacy risks to the people whose data was used to train models. Assessing the privacy risks of machine learning models is crucial to enabling knowledgeable decisions on whether to use, deploy, or share a model. A common approach to privacy risk assessment is to run one or more known attacks against the model and measure their success rate. We present a novel framework for running membership inference attacks against classification models. Our framework takes advantage of the ensemble method, generating many specialized attack models for different subsets of the data. We show that this approach achieves higher accuracy than either a single attack model or an attack model per class label, both on classical and language classification tasks.

Data Minimization for GDPR Compliance in Machine Learning Models

The EU General Data Protection Regulation (GDPR) mandates the principle of data minimization, which requires that only data necessary to fulfill a certain purpose be collected. However, it can often be difficult to determine the minimal amount of data required, especially in complex machine learning models such as neural networks. We present a first-of-a-kind method to reduce the amount of personal data needed to perform predictions with a machine learning model, by removing or generalizing some of the input features. Our method makes use of the knowledge encoded within the model to produce a generalization that has little to no impact on its accuracy. This enables the creators and users of machine learning models to acheive data minimization, in a provable manner.

Anonymizing Machine Learning Models

There is a known tension between the need to analyze personal data to drive business and privacy concerns. Many data protection regulations, including the EU General Data Protection Regulation (GDPR) and the California Consumer Protection Act (CCPA), set out strict restrictions and obligations on companies that collect or process personal data. Moreover, machine learning models themselves can be used to derive personal information, as demonstrated by recent membership and attribute inference attacks. Anonymized data, however, is exempt from data protection principles and obligations. Thus, models built on anonymized data are also exempt from any privacy obligations, in addition to providing better protection against such attacks on the training data. Learning on anonymized data typically results in a significant degradation in accuracy. We address this challenge by guiding our anonymization using the knowledge encoded within the model, and targeting it to minimize the impact on the model's accuracy, a process we call accuracy-guided anonymization. We demonstrate that by focusing on the model's accuracy rather than information loss, our method outperforms state of the art k-anonymity methods in terms of the achieved utility, in particular with high values of k and large numbers of quasi-identifiers. We also demonstrate that our approach achieves similar results in its ability to prevent membership inference attacks as alternative approaches based on differential privacy. This shows that model-guided anonymization can, in some cases, be a legitimate substitute for such methods, while averting some of their inherent drawbacks such as complexity, performance overhead and being fitted to specific model types. As opposed to methods that rely on adding noise during training, our approach does not rely on making any modifications to the training algorithm itself.

Reducing Risk of Model Inversion Using Privacy-Guided Training

Machine learning models often pose a threat to the privacy of individuals whose data is part of the training set. Several recent attacks have been able to infer sensitive information from trained models, including model inversion or attribute inference attacks. These attacks are able to reveal the values of certain sensitive features of individuals who participated in training the model. It has also been shown that several factors can contribute to an increased risk of model inversion, including feature influence. We observe that not all features necessarily share the same level of privacy or sensitivity. In many cases, certain features used to train a model are considered especially sensitive and therefore propitious candidates for inversion. We present a solution for countering model inversion attacks in tree-based models, by reducing the influence of sensitive features in these models. This is an avenue that has not yet been thoroughly investigated, with only very nascent previous attempts at using this as a countermeasure against attribute inference. Our work shows that, in many cases, it is possible to train a model in different ways, resulting in different influence levels of the various features, without necessarily harming the model's accuracy. We are able to utilize this fact to train models in a manner that reduces the model's reliance on the most sensitive features, while increasing the importance of less sensitive features. Our evaluation confirms that training models in this manner reduces the risk of inference for those features, as demonstrated through several black-box and white-box attacks.