Previous studies have revealed that artificial intelligence (AI) systems are vulnerable to adversarial attacks. Among them, model extraction attacks fool the target model by generating adversarial examples on a substitute model. The core of such an attack is training a substitute model as similar to the target model as possible, where the simulation process can be categorized in a data-dependent and data-free manner. Compared with the data-dependent method, the data-free one has been proven to be more practical in the real world since it trains the substitute model with synthesized data. However, the distribution of these fake data lacks diversity and cannot detect the decision boundary of the target model well, resulting in the dissatisfactory simulation effect. Besides, these data-free techniques need a vast number of queries to train the substitute model, increasing the time and computing consumption and the risk of exposure. To solve the aforementioned problems, in this paper, we propose a novel data-free model extraction method named SCME (Self-Contrastive Model Extraction), which considers both the inter- and intra-class diversity in synthesizing fake data. In addition, SCME introduces the Mixup operation to augment the fake data, which can explore the target model's decision boundary effectively and improve the simulating capacity. Extensive experiments show that the proposed method can yield diversified fake data. Moreover, our method has shown superiority in many different attack settings under the query-limited scenario, especially for untargeted attacks, the SCME outperforms SOTA methods by 11.43\% on average for five baseline datasets.
Model extraction attacks are designed to steal trained models with only query access, as is often provided through APIs that ML-as-a-Service providers offer. Machine Learning (ML) models are expensive to train, in part because data is hard to obtain, and a primary incentive for model extraction is to acquire a model while incurring less cost than training from scratch. Literature on model extraction commonly claims or presumes that the attacker is able to save on both data acquisition and labeling costs. We thoroughly evaluate this assumption and find that the attacker often does not. This is because current attacks implicitly rely on the adversary being able to sample from the victim model's data distribution. We thoroughly research factors influencing the success of model extraction. We discover that prior knowledge of the attacker, i.e., access to in-distribution data, dominates other factors like the attack policy the adversary follows to choose which queries to make to the victim model API. Our findings urge the community to redefine the adversarial goals of ME attacks as current evaluation methods misinterpret the ME performance.
Diffusion models showcased strong capabilities in image synthesis, being used in many computer vision tasks with great success. To this end, we propose to explore a new use case, namely to copy black-box classification models without having access to the original training data, the architecture, and the weights of the model, \ie~the model is only exposed through an inference API. More specifically, we can only observe the (soft or hard) labels for some image samples passed as input to the model. Furthermore, we consider an additional constraint limiting the number of model calls, mostly focusing our research on few-call model stealing. In order to solve the model extraction task given the applied restrictions, we propose the following framework. As training data, we create a synthetic data set (called proxy data set) by leveraging the ability of diffusion models to generate realistic and diverse images. Given a maximum number of allowed API calls, we pass the respective number of samples through the black-box model to collect labels. Finally, we distill the knowledge of the black-box teacher (attacked model) into a student model (copy of the attacked model), harnessing both labeled and unlabeled data generated by the diffusion model. We employ a novel active self-paced learning framework to make the most of the proxy data during distillation. Our empirical results on two data sets confirm the superiority of our framework over two state-of-the-art methods in the few-call model extraction scenario.
Business Process Management (BPM) is gaining increasing attention as it has the potential to cut costs while boosting output and quality. Business process document generation is a crucial stage in BPM. However, due to a shortage of datasets, data-driven deep learning techniques struggle to deliver the expected results. We propose an approach to transform Conditional Process Trees (CPTs) into Business Process Text Sketches (BPTSs) using Large Language Models (LLMs). The traditional prompting approach (Few-shot In-Context Learning) tries to get the correct answer in one go, and it can find the pattern of transforming simple CPTs into BPTSs, but for close-domain and CPTs with complex hierarchy, the traditional prompts perform weakly and with low correctness. We suggest using this technique to break down a difficult CPT into a number of basic CPTs and then solve each one in turn, drawing inspiration from the divide-and-conquer strategy. We chose 100 process trees with depths ranging from 2 to 5 at random, as well as CPTs with many nodes, many degrees of selection, and cyclic nesting. Experiments show that our method can achieve a correct rate of 93.42%, which is 45.17% better than traditional prompting methods. Our proposed method provides a solution for business process document generation in the absence of datasets, and secondly, it becomes potentially possible to provide a large number of datasets for the process model extraction (PME) domain.
Training a high-performance deep neural network requires large amounts of data and computational resources. Protecting the intellectual property (IP) and commercial ownership of a deep model is challenging yet increasingly crucial. A major stream of watermarking strategies implants verifiable backdoor triggers by poisoning training samples, but these are often unrealistic due to data privacy and safety concerns and are vulnerable to minor model changes such as fine-tuning. To overcome these challenges, we propose a safe and robust backdoor-based watermark injection technique that leverages the diverse knowledge from a single out-of-distribution (OoD) image, which serves as a secret key for IP verification. The independence of training data makes it agnostic to third-party promises of IP security. We induce robustness via random perturbation of model parameters during watermark injection to defend against common watermark removal attacks, including fine-tuning, pruning, and model extraction. Our experimental results demonstrate that the proposed watermarking approach is not only time- and sample-efficient without training data, but also robust against the watermark removal attacks above.
Model extraction emerges as a critical security threat with attack vectors exploiting both algorithmic and implementation-based approaches. The main goal of an attacker is to steal as much information as possible about a protected victim model, so that he can mimic it with a substitute model, even with a limited access to similar training data. Recently, physical attacks such as fault injection have shown worrying efficiency against the integrity and confidentiality of embedded models. We focus on embedded deep neural network models on 32-bit microcontrollers, a widespread family of hardware platforms in IoT, and the use of a standard fault injection strategy - Safe Error Attack (SEA) - to perform a model extraction attack with an adversary having a limited access to training data. Since the attack strongly depends on the input queries, we propose a black-box approach to craft a successful attack set. For a classical convolutional neural network, we successfully recover at least 90% of the most significant bits with about 1500 crafted inputs. These information enable to efficiently train a substitute model, with only 8% of the training dataset, that reaches high fidelity and near identical accuracy level than the victim model.
Due to the proliferation of malware, defenders are increasingly turning to automation and machine learning as part of the malware detection tool-chain. However, machine learning models are susceptible to adversarial attacks, requiring the testing of model and product robustness. Meanwhile, attackers also seek to automate malware generation and evasion of antivirus systems, and defenders try to gain insight into their methods. This work proposes a new algorithm that combines Malware Evasion and Model Extraction (MEME) attacks. MEME uses model-based reinforcement learning to adversarially modify Windows executable binary samples while simultaneously training a surrogate model with a high agreement with the target model to evade. To evaluate this method, we compare it with two state-of-the-art attacks in adversarial malware creation, using three well-known published models and one antivirus product as targets. Results show that MEME outperforms the state-of-the-art methods in terms of evasion capabilities in almost all cases, producing evasive malware with an evasion rate in the range of 32-73%. It also produces surrogate models with a prediction label agreement with the respective target models between 97-99%. The surrogate could be used to fine-tune and improve the evasion rate in the future.
A significant number of machine learning models are vulnerable to model extraction attacks, which focus on stealing the models by using specially curated queries against the target model. This task is well accomplished by using part of the training data or a surrogate dataset to train a new model that mimics a target model in a white-box environment. In pragmatic situations, however, the target models are trained on private datasets that are inaccessible to the adversary. The data-free model extraction technique replaces this problem when it comes to using queries artificially curated by a generator similar to that used in Generative Adversarial Nets. We propose for the first time, to the best of our knowledge, an adversary black box attack extending to a regression problem for predicting bounding box coordinates in object detection. As part of our study, we found that defining a loss function and using a novel generator setup is one of the key aspects in extracting the target model. We find that the proposed model extraction method achieves significant results by using reasonable queries. The discovery of this object detection vulnerability will support future prospects for securing such models.
DNN accelerators have been widely deployed in many scenarios to speed up the inference process and reduce the energy consumption. One big concern about the usage of the accelerators is the confidentiality of the deployed models: model inference execution on the accelerators could leak side-channel information, which enables an adversary to preciously recover the model details. Such model extraction attacks can not only compromise the intellectual property of DNN models, but also facilitate some adversarial attacks. Although previous works have demonstrated a number of side-channel techniques to extract models from DNN accelerators, they are not practical for two reasons. (1) They only target simplified accelerator implementations, which have limited practicality in the real world. (2) They require heavy human analysis and domain knowledge. To overcome these limitations, this paper presents Mercury, the first automated remote side-channel attack against the off-the-shelf Nvidia DNN accelerator. The key insight of Mercury is to model the side-channel extraction process as a sequence-to-sequence problem. The adversary can leverage a time-to-digital converter (TDC) to remotely collect the power trace of the target model's inference. Then he uses a learning model to automatically recover the architecture details of the victim model from the power trace without any prior knowledge. The adversary can further use the attention mechanism to localize the leakage points that contribute most to the attack. Evaluation results indicate that Mercury can keep the error rate of model extraction below 1%.
This research focuses on the evolving dynamics of the power grid, where traditional synchronous generators are being replaced by non-synchronous power electronic converter (PEC)-interfaced renewable energy sources. The non-linear dynamics must be accurately modeled to ensure the stability of future converter-dominated power systems (CDPS). However, obtaining comprehensive dynamic models becomes more complex and computationally intensive as the system grows. This study proposes a scalable and automated data-driven partitioned modeling framework for CDPS dynamics. The method constructs reduced-ordered dynamic linear transfer function models using input-output measurements from a PEC switching model. Validation experiments were conducted on single-house and multi-house scenarios, demonstrating high accuracy (over 97%) and significant computational speed improvements (6.5 times faster) compared to comprehensive models. This framework and modeling approach offers valuable insights for efficient analysis of power system dynamics, aiding in planning, operation, and dispatch.
Extracting workflow nets from textual descriptions can be used to simplify guidelines or formalize textual descriptions of formal processes like business processes and algorithms. The task of manually extracting processes, however, requires domain expertise and effort. While automatic process model extraction is desirable, annotating texts with formalized process models is expensive. Therefore, there are only a few machine-learning-based extraction approaches. Rule-based approaches, in turn, require domain specificity to work well and can rarely distinguish relevant and irrelevant information in textual descriptions. In this paper, we present GUIDO, a hybrid approach to the process model extraction task that first, classifies sentences regarding their relevance to the process model, using a BERT-based sentence classifier, and second, extracts a process model from the sentences classified as relevant, using dependency parsing. The presented approach achieves significantly better results than a pure rule-based approach. GUIDO achieves an average behavioral similarity score of $0.93$. Still, in comparison to purely machine-learning-based approaches, the annotation costs stay low.
Model-serving systems have become increasingly popular, especially in real-time web applications. In such systems, users send queries to the server and specify the desired performance metrics (e.g., desired accuracy, latency). The server maintains a set of models (model zoo) in the back-end and serves the queries based on the specified metrics. This paper examines the security, specifically robustness against model extraction attacks, of such systems. Existing black-box attacks assume a single model can be repeatedly selected for serving inference requests. Modern inference serving systems break this assumption. Thus, they cannot be directly applied to extract a victim model, as models are hidden behind a layer of abstraction exposed by the serving system. An attacker can no longer identify which model she is interacting with. To this end, we first propose a query-efficient fingerprinting algorithm to enable the attacker to trigger any desired model consistently. We show that by using our fingerprinting algorithm, model extraction can have fidelity and accuracy scores within $1\%$ of the scores obtained when attacking a single, explicitly specified model, as well as up to $14.6\%$ gain in accuracy and up to $7.7\%$ gain in fidelity compared to the naive attack. Second, we counter the proposed attack with a noise-based defense mechanism that thwarts fingerprinting by adding noise to the specified performance metrics. The proposed defense strategy reduces the attack's accuracy and fidelity by up to $9.8\%$ and $4.8\%$, respectively (on medium-sized model extraction). Third, we show that the proposed defense induces a fundamental trade-off between the level of protection and system goodput, achieving configurable and significant victim model extraction protection while maintaining acceptable goodput ($>80\%$). We implement the proposed defense in a real system with plans to open source.
Recurrent Neural Networks (RNNs) have achieved tremendous success in processing sequential data, yet understanding and analyzing their behaviours remains a significant challenge. To this end, many efforts have been made to extract finite automata from RNNs, which are more amenable for analysis and explanation. However, existing approaches like exact learning and compositional approaches for model extraction have limitations in either scalability or precision. In this paper, we propose a novel framework of Weighted Finite Automata (WFA) extraction and explanation to tackle the limitations for natural language tasks. First, to address the transition sparsity and context loss problems we identified in WFA extraction for natural language tasks, we propose an empirical method to complement missing rules in the transition diagram, and adjust transition matrices to enhance the context-awareness of the WFA. We also propose two data augmentation tactics to track more dynamic behaviours of RNN, which further allows us to improve the extraction precision. Based on the extracted model, we propose an explanation method for RNNs including a word embedding method -- Transition Matrix Embeddings (TME) and TME-based task oriented explanation for the target RNN. Our evaluation demonstrates the advantage of our method in extraction precision than existing approaches, and the effectiveness of TME-based explanation method in applications to pretraining and adversarial example generation.
Machine Learning as a Service (MLaaS) platforms have gained popularity due to their accessibility, cost-efficiency, scalability, and rapid development capabilities. However, recent research has highlighted the vulnerability of cloud-based models in MLaaS to model extraction attacks. In this paper, we introduce FDINET, a novel defense mechanism that leverages the feature distribution of deep neural network (DNN) models. Concretely, by analyzing the feature distribution from the adversary's queries, we reveal that the feature distribution of these queries deviates from that of the model's training set. Based on this key observation, we propose Feature Distortion Index (FDI), a metric designed to quantitatively measure the feature distribution deviation of received queries. The proposed FDINET utilizes FDI to train a binary detector and exploits FDI similarity to identify colluding adversaries from distributed extraction attacks. We conduct extensive experiments to evaluate FDINET against six state-of-the-art extraction attacks on four benchmark datasets and four popular model architectures. Empirical results demonstrate the following findings FDINET proves to be highly effective in detecting model extraction, achieving a 100% detection accuracy on DFME and DaST. FDINET is highly efficient, using just 50 queries to raise an extraction alarm with an average confidence of 96.08% for GTSRB. FDINET exhibits the capability to identify colluding adversaries with an accuracy exceeding 91%. Additionally, it demonstrates the ability to detect two types of adaptive attacks.
Generative adversarial networks (GANs) have shown remarkable success in image synthesis, making GAN models themselves commercially valuable to legitimate model owners. Therefore, it is critical to technically protect the intellectual property of GANs. Prior works need to tamper with the training set or training process, and they are not robust to emerging model extraction attacks. In this paper, we propose a new ownership protection method based on the common characteristics of a target model and its stolen models. Our method can be directly applicable to all well-trained GANs as it does not require retraining target models. Extensive experimental results show that our new method can achieve the best protection performance, compared to the state-of-the-art methods. Finally, we demonstrate the effectiveness of our method with respect to the number of generations of model extraction attacks, the number of generated samples, different datasets, as well as adaptive attacks.
Deep neural network (DNN) models have become a critical asset of the model owner as training them requires a large amount of resource (i.e. labeled data). Therefore, many fingerprinting schemes have been proposed to safeguard the intellectual property (IP) of the model owner against model extraction and illegal redistribution. However, previous schemes adopt unnatural images as the fingerprint, such as adversarial examples and noisy images, which can be easily perceived and rejected by the adversary. In this paper, we propose NaturalFinger which generates natural fingerprint with generative adversarial networks (GANs). Besides, our proposed NaturalFinger fingerprints the decision difference areas rather than the decision boundary, which is more robust. The application of GAN not only allows us to generate more imperceptible samples, but also enables us to generate unrestricted samples to explore the decision boundary.To demonstrate the effectiveness of our fingerprint approach, we evaluate our approach against four model modification attacks including adversarial training and two model extraction attacks. Experiments show that our approach achieves 0.91 ARUC value on the FingerBench dataset (154 models), exceeding the optimal baseline (MetaV) over 17\%.
Large language models (LLMs) have demonstrated powerful capabilities in both text understanding and generation. Companies have begun to offer Embedding as a Service (EaaS) based on these LLMs, which can benefit various natural language processing (NLP) tasks for customers. However, previous studies have shown that EaaS is vulnerable to model extraction attacks, which can cause significant losses for the owners of LLMs, as training these models is extremely expensive. To protect the copyright of LLMs for EaaS, we propose an Embedding Watermark method called EmbMarker that implants backdoors on embeddings. Our method selects a group of moderate-frequency words from a general text corpus to form a trigger set, then selects a target embedding as the watermark, and inserts it into the embeddings of texts containing trigger words as the backdoor. The weight of insertion is proportional to the number of trigger words included in the text. This allows the watermark backdoor to be effectively transferred to EaaS-stealer's model for copyright verification while minimizing the adverse impact on the original embeddings' utility. Our extensive experiments on various datasets show that our method can effectively protect the copyright of EaaS models without compromising service quality.
We introduce the problem of model-extraction attacks in cyber-physical systems in which an attacker attempts to estimate (or extract) the feedback controller of the system. Extracting (or estimating) the controller provides an unmatched edge to attackers since it allows them to predict the future control actions of the system and plan their attack accordingly. Hence, it is important to understand the ability of the attackers to perform such an attack. In this paper, we focus on the setting when a Deep Neural Network (DNN) controller is trained using Reinforcement Learning (RL) algorithms and is used to control a stochastic system. We play the role of the attacker that aims to estimate such an unknown DNN controller, and we propose a two-phase algorithm. In the first phase, also called the offline phase, the attacker uses side-channel information about the RL-reward function and the system dynamics to identify a set of candidate estimates of the unknown DNN. In the second phase, also called the online phase, the attacker observes the behavior of the unknown DNN and uses these observations to shortlist the set of final policy estimates. We provide theoretical analysis of the error between the unknown DNN and the estimated one. We also provide numerical results showing the effectiveness of the proposed algorithm.
Graph neural networks (GNNs) have emerged as a state-of-the-art approach to model and draw inferences from large scale graph-structured data in various application settings such as social networking. The primary goal of a GNN is to learn an embedding for each graph node in a dataset that encodes both the node features and the local graph structure around the node. Embeddings generated by a GNN for a graph node are unique to that GNN. Prior work has shown that GNNs are prone to model extraction attacks. Model extraction attacks and defenses have been explored extensively in other non-graph settings. While detecting or preventing model extraction appears to be difficult, deterring them via effective ownership verification techniques offer a potential defense. In non-graph settings, fingerprinting models, or the data used to build them, have shown to be a promising approach toward ownership verification. We present GrOVe, a state-of-the-art GNN model fingerprinting scheme that, given a target model and a suspect model, can reliably determine if the suspect model was trained independently of the target model or if it is a surrogate of the target model obtained via model extraction. We show that GrOVe can distinguish between surrogate and independent models even when the independent model uses the same training dataset and architecture as the original target model. Using six benchmark datasets and three model architectures, we show that consistently achieves low false-positive and false-negative rates. We demonstrate that is robust against known fingerprint evasion techniques while remaining computationally efficient.
Deep Neural Networks (DNNs) have become ubiquitous due to their performance on prediction and classification problems. However, they face a variety of threats as their usage spreads. Model extraction attacks, which steal DNNs, endanger intellectual property, data privacy, and security. Previous research has shown that system-level side-channels can be used to leak the architecture of a victim DNN, exacerbating these risks. We propose two DNN architecture extraction techniques catering to various threat models. The first technique uses a malicious, dynamically linked version of PyTorch to expose a victim DNN architecture through the PyTorch profiler. The second, called EZClone, exploits aggregate (rather than time-series) GPU profiles as a side-channel to predict DNN architecture, employing a simple approach and assuming little adversary capability as compared to previous work. We investigate the effectiveness of EZClone when minimizing the complexity of the attack, when applied to pruned models, and when applied across GPUs. We find that EZClone correctly predicts DNN architectures for the entire set of PyTorch vision architectures with 100% accuracy. No other work has shown this degree of architecture prediction accuracy with the same adversarial constraints or using aggregate side-channel information. Prior work has shown that, once a DNN has been successfully cloned, further attacks such as model evasion or model inversion can be accelerated significantly.