Sebastian Szyller

Mansi Phute, Alec Helbling, Matthew Hull et al. · gatech

Large language models (LLMs) are popular for high-quality text generation but can produce harmful content, even when aligned with human values through reinforcement learning. Adversarial prompts can bypass their safety measures. We propose LLM Self Defense, a simple approach to defend against these attacks by having an LLM screen the induced responses. Our method does not require any fine-tuning, input preprocessing, or iterative output generation. Instead, we incorporate the generated content into a pre-defined prompt and employ another instance of an LLM to analyze the text and predict whether it is harmful. We test LLM Self Defense on GPT 3.5 and Llama 2, two of the current most prominent LLMs against various types of attacks, such as forcefully inducing affirmative responses to prompts and prompt engineering attacks. Notably, LLM Self Defense succeeds in reducing the attack success rate to virtually 0 using both GPT 3.5 and Llama 2. The code is publicly available at https://github.com/poloclub/llm-self-defense

Sebastian Szyller, Rui Zhang, Jian Liu et al.

Machine learning (ML) models are costly to train as they can require a significant amount of data, computational resources and technical expertise. Thus, they constitute valuable intellectual property that needs protection from adversaries wanting to steal them. Ownership verification techniques allow the victims of model stealing attacks to demonstrate that a suspect model was in fact stolen from theirs. Although a number of ownership verification techniques based on watermarking or fingerprinting have been proposed, most of them fall short either in terms of security guarantees (well-equipped adversaries can evade verification) or computational cost. A fingerprinting technique, Dataset Inference (DI), has been shown to offer better robustness and efficiency than prior methods. The authors of DI provided a correctness proof for linear (suspect) models. However, in a subspace of the same setting, we prove that DI suffers from high false positives (FPs) -- it can incorrectly identify an independent model trained with non-overlapping data from the same distribution as stolen. We further prove that DI also triggers FPs in realistic, non-linear suspect models. We then confirm empirically that DI in the black-box setting leads to FPs, with high confidence. Second, we show that DI also suffers from false negatives (FNs) -- an adversary can fool DI (at the cost of incurring some accuracy loss) by regularising a stolen model's decision boundaries using adversarial training, thereby leading to an FN. To this end, we demonstrate that black-box DI fails to identify a model adversarially trained from a stolen dataset -- the setting where DI is the hardest to evade. Finally, we discuss the implications of our findings, the viability of fingerprinting-based ownership verification in general, and suggest directions for future work.

Jian Liu, Rui Zhang, Sebastian Szyller et al.

Deep neural network (DNN) models are valuable intellectual property of model owners, constituting a competitive advantage. Therefore, it is crucial to develop techniques to protect against model theft. Model ownership resolution (MOR) is a class of techniques that can deter model theft. A MOR scheme enables an accuser to assert an ownership claim for a suspect model by presenting evidence, such as a watermark or fingerprint, to show that the suspect model was stolen or derived from a source model owned by the accuser. Most of the existing MOR schemes prioritize robustness against malicious suspects, ensuring that the accuser will win if the suspect model is indeed a stolen model. In this paper, we show that common MOR schemes in the literature are vulnerable to a different, equally important but insufficiently explored, robustness concern: a malicious accuser. We show how malicious accusers can successfully make false claims against independent suspect models that were not stolen. Our core idea is that a malicious accuser can deviate (without detection) from the specified MOR process by finding (transferable) adversarial examples that successfully serve as evidence against independent suspect models. To this end, we first generalize the procedures of common MOR schemes and show that, under this generalization, defending against false claims is as challenging as preventing (transferable) adversarial examples. Via systematic empirical evaluation, we show that our false claim attacks always succeed in the MOR schemes that follow our generalization, including in a real-world model: Amazon's Rekognition API.

Sebastian Szyller, N. Asokan

Nowadays, systems based on machine learning (ML) are widely used in different domains. Given their popularity, ML models have become targets for various attacks. As a result, research at the intersection of security/privacy and ML has flourished. Typically such work has focused on individual types of security/privacy concerns and mitigations thereof. However, in real-life deployments, an ML model will need to be protected against several concerns simultaneously. A protection mechanism optimal for one security or privacy concern may interact negatively with mechanisms intended to address other concerns. Despite its practical relevance, the potential for such conflicts has not been studied adequately. We first provide a framework for analyzing such "conflicting interactions". We then focus on systematically analyzing pairwise interactions between protection mechanisms for one concern, model and data ownership verification, with two other classes of ML protection mechanisms: differentially private training, and robustness against model evasion. We find that several pairwise interactions result in conflicts. We explore potential approaches for avoiding such conflicts. First, we study the effect of hyperparameter relaxations, finding that there is no sweet spot balancing the performance of both protection mechanisms. Second, we explore if modifying one type of protection mechanism (ownership verification) so as to decouple it from factors that may be impacted by a conflicting mechanism (differentially private training or robustness to model evasion) can avoid conflict. We show that this approach can avoid the conflict between ownership verification mechanisms when combined with differentially private training, but has no effect on robustness to model evasion. Finally, we identify the gaps in the landscape of studying interactions between other types of ML protection mechanisms.

Marcin Spoczynski, Marcela S. Melara, Sebastian Szyller

The rapid adoption of open source machine learning (ML) datasets and models exposes today's AI applications to critical risks like data poisoning and supply chain attacks across the ML lifecycle. With growing regulatory pressure to address these issues through greater transparency, ML model vendors face challenges balancing these requirements against confidentiality for data and intellectual property needs. We propose Atlas, a framework that enables fully attestable ML pipelines. Atlas leverages open specifications for data and software supply chain provenance to collect verifiable records of model artifact authenticity and end-to-end lineage metadata. Atlas combines trusted hardware and transparency logs to enhance metadata integrity, preserve data confidentiality, and limit unauthorized access during ML pipeline operations, from training through deployment. Our prototype implementation of Atlas integrates several open-source tools to build an ML lifecycle transparency system, and assess the practicality of Atlas through two case study ML pipelines.

Haokun Chen, Sebastian Szyller, Weilin Xu et al.

Large language models (LLMs) are trained using massive datasets, which often contain undesirable content such as harmful texts, personal information, and copyrighted material. To address this, machine unlearning aims to remove information from trained models. Recent work has shown that soft token attacks (STA) can successfully extract unlearned information from LLMs, but in this work we show that STAs can be an inadequate tool for auditing unlearning. Using common benchmarks such as Who Is Harry Potter? and TOFU, we demonstrate that in a strong auditor setting such attacks can elicit any information from the LLM, regardless of the deployed unlearning algorithm or whether the queried content was originally present in the training corpus. We further show that STA with just a few soft tokens (1-10) can elicit random strings over 400 characters long, indicating that STAs must be used carefully to effectively audit unlearning. Example code can be found at: https://github.com/IntelLabs/LLMart/tree/main/examples/unlearning

Vasisht Duddu, Sebastian Szyller, N. Asokan

Machine learning (ML) models cannot neglect risks to security, privacy, and fairness. Several defenses have been proposed to mitigate such risks. When a defense is effective in mitigating one risk, it may correspond to increased or decreased susceptibility to other risks. Existing research lacks an effective framework to recognize and explain these unintended interactions. We present such a framework, based on the conjecture that overfitting and memorization underlie unintended interactions. We survey existing literature on unintended interactions, accommodating them within our framework. We use our framework to conjecture on two previously unexplored interactions, and empirically validate our conjectures.

Asim Waheed, Vasisht Duddu, Rui Zhang et al.

Machine learning (ML) models are susceptible to various risks to security, privacy, and fairness. Most defenses are designed to protect against each risk individually (intended interactions) but can inadvertently affect susceptibility to other unrelated risks (unintended interactions). We introduce Amulet, the first Python library for evaluating both intended and unintended interactions among ML defenses and risks. Amulet is comprehensive by including representative attacks, defenses, and metrics; extensible to new modules due to its modular design; consistent with a user-friendly API template for inputs and outputs; and applicable for evaluating novel interactions. By satisfying all four properties, Amulet offers a unified foundation for studying how defenses interact, enabling the first systematic evaluation of unintended interactions across multiple risks.

Weilin Xu, Sebastian Szyller, Cory Cornelius et al.

Adversarial examples in the digital domain against deep learning-based computer vision models allow for perturbations that are imperceptible to human eyes. However, producing similar adversarial examples in the physical world has been difficult due to the non-differentiable image distortion functions in visual sensing systems. The existing algorithms for generating physically realizable adversarial examples often loosen their definition of adversarial examples by allowing unbounded perturbations, resulting in obvious or even strange visual patterns. In this work, we make adversarial examples imperceptible in the physical world using a straight-through estimator (STE, a.k.a. BPDA). We employ STE to overcome the non-differentiability -- applying exact, non-differentiable distortions in the forward pass of the backpropagation step, and using the identity function in the backward pass. Our differentiable rendering extension to STE also enables imperceptible adversarial patches in the physical world. Using printout photos, and experiments in the CARLA simulator, we show that STE enables fast generation of $\ell_\infty$ bounded adversarial examples despite the non-differentiable distortions. To the best of our knowledge, this is the first work demonstrating imperceptible adversarial examples bounded by small $\ell_\infty$ norms in the physical world that force zero classification accuracy in the global perturbation threat model and cause near-zero ($4.22\%$) AP50 in object detection in the patch perturbation threat model. We urge the community to re-evaluate the threat of adversarial examples in the physical world.

Vasisht Duddu, Sebastian Szyller, N. Asokan

Data used to train machine learning (ML) models can be sensitive. Membership inference attacks (MIAs), attempting to determine whether a particular data record was used to train an ML model, risk violating membership privacy. ML model builders need a principled definition of a metric to quantify the membership privacy risk of (a) individual training data records, (b) computed independently of specific MIAs, (c) which assesses susceptibility to different MIAs, (d) can be used for different applications, and (e) efficiently. None of the prior membership privacy risk metrics simultaneously meet all these requirements. We present SHAPr, a membership privacy metric based on Shapley values which is a leave-one-out (LOO) technique, originally intended to measure the contribution of a training data record on model utility. We conjecture that contribution to model utility can act as a proxy for memorization, and hence represent membership privacy risk. Using ten benchmark datasets, we show that SHAPr is indeed effective in estimating susceptibility of training data records to MIAs. We also show that, unlike prior work, SHAPr is significantly better in estimating susceptibility to newer, and more effective MIA. We apply SHAPr to evaluate the efficacy of several defenses against MIAs: using regularization and removing high risk training data records. Moreover, SHAPr is versatile: it can be used for estimating vulnerability of different subgroups to MIAs, and inherits applications of Shapley values (e.g., data valuation). We show that SHAPr has an acceptable computational cost (compared to naive LOO), varying from a few minutes for the smallest dataset to ~92 minutes for the largest dataset.

Sebastian Szyller, Vasisht Duddu, Tommi Gröndahl et al.

Machine learning models are typically made available to potential client users via inference APIs. Model extraction attacks occur when a malicious client uses information gleaned from queries to the inference API of a victim model $F_V$ to build a surrogate model $F_A$ with comparable functionality. Recent research has shown successful model extraction of image classification, and natural language processing models. In this paper, we show the first model extraction attack against real-world generative adversarial network (GAN) image translation models. We present a framework for conducting such attacks, and show that an adversary can successfully extract functional surrogate models by querying $F_V$ using data from the same domain as the training data for $F_V$. The adversary need not know $F_V$'s architecture or any other information about it beyond its intended task. We evaluate the effectiveness of our attacks using three different instances of two popular categories of image translation: (1) Selfie-to-Anime and (2) Monet-to-Photo (image style transfer), and (3) Super-Resolution (super resolution). Using standard performance metrics for GANs, we show that our attacks are effective. Furthermore, we conducted a large scale (125 participants) user study on Selfie-to-Anime and Monet-to-Photo to show that human perception of the images produced by $F_V$ and $F_A$ can be considered equivalent, within an equivalence bound of Cohen's d = 0.3. Finally, we show that existing defenses against model extraction attacks (watermarking, adversarial examples, poisoning) do not extend to image translation models.

Buse Gul Atli, Sebastian Szyller, Mika Juuti et al.

Recently, machine learning (ML) has introduced advanced solutions to many domains. Since ML models provide business advantage to model owners, protecting intellectual property of ML models has emerged as an important consideration. Confidentiality of ML models can be protected by exposing them to clients only via prediction APIs. However, model extraction attacks can steal the functionality of ML models using the information leaked to clients through the results returned via the API. In this work, we question whether model extraction is a serious threat to complex, real-life ML models. We evaluate the current state-of-the-art model extraction attack (Knockoff nets) against complex models. We reproduce and confirm the results in the original paper. But we also show that the performance of this attack can be limited by several factors, including ML model architecture and the granularity of API response. Furthermore, we introduce a defense based on distinguishing queries used for Knockoff nets from benign queries. Despite the limitations of the Knockoff nets, we show that a more realistic adversary can effectively steal complex ML models and evade known defenses.

Samuel Marchal, Sebastian Szyller

Online retail, eCommerce, frequently falls victim to fraud conducted by malicious customers (fraudsters) who obtain goods or services through deception. Fraud coordinated by groups of professional fraudsters that place several fraudulent orders to maximize their gain is referred to as organized fraud. Existing approaches to fraud detection typically analyze orders in isolation and they are not effective at identifying groups of fraudulent orders linked to organized fraud. These also wrongly identify many legitimate orders as fraud, which hinders their usage for automated fraud cancellation. We introduce a novel solution to detect organized fraud by analyzing orders in bulk. Our approach is based on clustering and aims to group together fraudulent orders placed by the same group of fraudsters. It selectively uses two existing techniques, agglomerative clustering and sampling to recursively group orders into small clusters in a reasonable amount of time. We assess our clustering technique on real-world orders placed on the Zalando website, the largest online apparel retailer in Europe1. Our clustering processes 100,000s of orders in a few hours and groups 35-45% of fraudulent orders together. We propose a simple technique built on top of our clustering that detects 26.2% of fraud while raising false alarms for only 0.1% of legitimate orders.

Sebastian Szyller, Buse Gul Atli, Samuel Marchal et al.

Training machine learning (ML) models is expensive in terms of computational power, amounts of labeled data and human expertise. Thus, ML models constitute intellectual property (IP) and business value for their owners. Embedding digital watermarks during model training allows a model owner to later identify their models in case of theft or misuse. However, model functionality can also be stolen via model extraction, where an adversary trains a surrogate model using results returned from a prediction API of the original model. Recent work has shown that model extraction is a realistic threat. Existing watermarking schemes are ineffective against IP theft via model extraction since it is the adversary who trains the surrogate model. In this paper, we introduce DAWN (Dynamic Adversarial Watermarking of Neural Networks), the first approach to use watermarking to deter model extraction IP theft. Unlike prior watermarking schemes, DAWN does not impose changes to the training process but it operates at the prediction API of the protected model, by dynamically changing the responses for a small subset of queries (e.g., <0.5%) from API clients. This set is a watermark that will be embedded in case a client uses its queries to train a surrogate model. We show that DAWN is resilient against two state-of-the-art model extraction attacks, effectively watermarking all extracted surrogate models, allowing model owners to reliably demonstrate ownership (with confidence $>1- 2^{-64}$), incurring negligible loss of prediction accuracy (0.03-0.5%).

Mika Juuti, Sebastian Szyller, Samuel Marchal et al.

Machine learning (ML) applications are increasingly prevalent. Protecting the confidentiality of ML models becomes paramount for two reasons: (a) a model can be a business advantage to its owner, and (b) an adversary may use a stolen model to find transferable adversarial examples that can evade classification by the original model. Access to the model can be restricted to be only via well-defined prediction APIs. Nevertheless, prediction APIs still provide enough information to allow an adversary to mount model extraction attacks by sending repeated queries via the prediction API. In this paper, we describe new model extraction attacks using novel approaches for generating synthetic queries, and optimizing training hyperparameters. Our attacks outperform state-of-the-art model extraction in terms of transferability of both targeted and non-targeted adversarial examples (up to +29-44 percentage points, pp), and prediction accuracy (up to +46 pp) on two datasets. We provide take-aways on how to perform effective model extraction attacks. We then propose PRADA, the first step towards generic and effective detection of DNN model extraction attacks. It analyzes the distribution of consecutive API queries and raises an alarm when this distribution deviates from benign behavior. We show that PRADA can detect all prior model extraction attacks with no false positives.