Biblio

Trojan attacks threaten deep neural networks (DNNs) by poisoning them to behave normally on most samples, yet to produce manipulated results for inputs attached with a particular trigger. Several works attempt to detect whether a given DNN has been injected with a specific trigger during the training. In a parallel line of research, the lottery ticket hypothesis reveals the existence of sparse sub-networks which are capable of reaching competitive performance as the dense network after independent training. Connecting these two dots, we investigate the problem of Trojan DNN detection from the brand new lens of sparsity, even when no clean training data is available. Our crucial observation is that the Trojan features are significantly more stable to network pruning than benign features. Leveraging that, we propose a novel Trojan network detection regime: first locating a “winning Trojan lottery ticket” which preserves nearly full Trojan information yet only chance-level performance on clean inputs; then recovering the trigger embedded in this already isolated sub-network. Extensive experiments on various datasets, i.e., CIFAR-10, CIFAR-100, and ImageNet, with different network architectures, i.e., VGG-16, ResNet-18, ResNet-20s, and DenseNet-100 demonstrate the effectiveness of our proposal. Codes are available at https://github.com/VITA-Group/Backdoor-LTH.

Existing defense strategies against adversarial attacks (AAs) on AI/ML are primarily focused on examining the input data streams using a wide variety of filtering techniques. For instance, input filters are used to remove noisy, misleading, and out-of-class inputs along with a variety of attacks on learning systems. However, a single filter may not be able to detect all types of AAs. To address this issue, in the current work, we propose a robust, transferable, distribution-independent, and cross-domain supported framework for selecting Adaptive Filter Ensembles (AFEs) to minimize the impact of data poisoning on learning systems. The optimal filter ensembles are determined through a Multi-Objective Bi-Level Programming Problem (MOBLPP) that provides a subset of diverse filter sequences, each exhibiting fair detection accuracy. The proposed framework of AFE is trained to model the pristine data distribution to identify the corrupted inputs and converges to the optimal AFE without vanishing gradients and mode collapses irrespective of input data distributions. We presented preliminary experiments to show the proposed defense outperforms the existing defenses in terms of robustness and accuracy.

Machine Learning (ML) and Artificial Intelligence (AI) techniques are widely adopted in the telecommunication industry, especially to automate beyond 5G networks. Federated Learning (FL) recently emerged as a distributed ML approach that enables localized model training to keep data decentralized to ensure data privacy. In this paper, we identify the applicability of FL for securing future networks and its limitations due to the vulnerability to poisoning attacks. First, we investigate the shortcomings of state-of-the-art security algorithms for FL and perform an attack to circumvent FoolsGold algorithm, which is known as one of the most promising defense techniques currently available. The attack is launched with the addition of intelligent noise at the poisonous model updates. Then we propose a more sophisticated defense strategy, a threshold-based clustering mechanism to complement FoolsGold. Moreover, we provide a comprehensive analysis of the impact of the attack scenario and the performance of the defense mechanism.

Federated learning (FL) has emerged as a promising paradigm for distributed training of machine learning models. In FL, several participants train a global model collaboratively by only sharing model parameter updates while keeping their training data local. However, FL was recently shown to be vulnerable to data poisoning attacks, in which malicious participants send parameter updates derived from poisoned training data. In this paper, we focus on defending against targeted data poisoning attacks, where the attacker’s goal is to make the model misbehave for a small subset of classes while the rest of the model is relatively unaffected. To defend against such attacks, we first propose a method called MAPPS for separating malicious updates from benign ones. Using MAPPS, we propose three methods for attack detection: MAPPS + X-Means, MAPPS + VAT, and their Ensemble. Then, we propose an attack mitigation approach in which a "clean" model (i.e., a model that is not negatively impacted by an attack) can be trained despite the existence of a poisoning attempt. We empirically evaluate all of our methods using popular image classification datasets. Results show that we can achieve \textgreater 95% true positive rates while incurring only \textless 2% false positive rate. Furthermore, the clean models that are trained using our proposed methods have accuracy comparable to models trained in an attack-free scenario.

In recent years, the security of AI systems has drawn increasing research attention, especially in the medical imaging realm. To develop a secure medical image analysis (MIA) system, it is a must to study possible backdoor attacks (BAs), which can embed hidden malicious behaviors into the system. However, designing a unified BA method that can be applied to various MIA systems is challenging due to the diversity of imaging modalities (e.g., X-Ray, CT, and MRI) and analysis tasks (e.g., classification, detection, and segmentation). Most existing BA methods are designed to attack natural image classification models, which apply spatial triggers to training images and inevitably corrupt the semantics of poisoned pixels, leading to the failures of attacking dense prediction models. To address this issue, we propose a novel Frequency-Injection based Backdoor Attack method (FIBA) that is capable of delivering attacks in various MIA tasks. Specifically, FIBA leverages a trigger function in the frequency domain that can inject the low-frequency information of a trigger image into the poisoned image by linearly combining the spectral amplitude of both images. Since it preserves the semantics of the poisoned image pixels, FIBA can perform attacks on both classification and dense prediction models. Experiments on three benchmarks in MIA (i.e., ISIC-2019 [4] for skin lesion classification, KiTS-19 [17] for kidney tumor segmentation, and EAD-2019 [1] for endoscopic artifact detection), validate the effectiveness of FIBA and its superiority over stateof-the-art methods in attacking MIA models and bypassing backdoor defense. Source code will be available at code.

With the widespread deployment of data-driven services, the demand for data volumes continues to grow. At present, many applications lack reliable human supervision in the process of data collection, which makes the collected data contain low-quality data or even malicious data. This low-quality or malicious data make AI systems potentially face much security challenges. One of the main security threats in the training phase of machine learning is data poisoning attacks, which compromise model integrity by contaminating training data to make the resulting model skewed or unusable. This paper reviews the relevant researches on data poisoning attacks in various task environments: first, the classification of attacks is summarized, then the defense methods of data poisoning attacks are sorted out, and finally, the possible research directions in the prospect.

Recent trends in the convergence of edge computing and artificial intelligence (AI) have led to a new paradigm of “edge intelligence”, which are more vulnerable to attack such as data and model poisoning and evasion of attacks. This paper proposes a white-box poisoning attack against online regression model for edge intelligence environment, which aim to prepare the protection methods in the future. Firstly, the new method selects data points from original stream with maximum loss by two selection strategies; Secondly, it pollutes these points with gradient ascent strategy. At last, it injects polluted points into original stream being sent to target model to complete the attack process. We extensively evaluate our proposed attack on open dataset, the results of which demonstrate the effectiveness of the novel attack method and the real implications of poisoning attack in a case study electric energy prediction application.

Graph-based Semi-Supervised Learning (GSSL) is a practical solution to learn from a limited amount of labelled data together with a vast amount of unlabelled data. However, due to their reliance on the known labels to infer the unknown labels, these algorithms are sensitive to data quality. It is therefore essential to study the potential threats related to the labelled data, more specifically, label poisoning. In this paper, we propose a novel data poisoning method which efficiently approximates the result of label inference to identify the inputs which, if poisoned, would produce the highest number of incorrectly inferred labels. We extensively evaluate our approach on three classification problems under 24 different experimental settings each. Compared to the state of the art, our influence-driven attack produces an average increase of error rate 50% higher, while being faster by multiple orders of magnitude. Moreover, our method can inform engineers of inputs that deserve investigation (relabelling them) before training the learning model. We show that relabelling one-third of the poisoned inputs (selected based on their influence) reduces the poisoning effect by 50%. ACM Reference Format: Adriano Franci, Maxime Cordy, Martin Gubri, Mike Papadakis, and Yves Le Traon. 2022. Influence-Driven Data Poisoning in Graph-Based Semi-Supervised Classifiers. In 1st Conference on AI Engineering - Software Engineering for AI (CAIN’22), May 16–24, 2022, Pittsburgh, PA, USA. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3522664.3528606

Speaker verification has been widely and successfully adopted in many mission-critical areas for user identification. The training of speaker verification requires a large amount of data, therefore users usually need to adopt third-party data (e.g., data from the Internet or third-party data company). This raises the question of whether adopting untrusted third-party data can pose a security threat. In this paper, we demonstrate that it is possible to inject the hidden backdoor for infecting speaker verification models by poisoning the training data. Specifically, we design a clustering-based attack scheme where poisoned samples from different clusters will contain different triggers (i.e., pre-defined utterances), based on our understanding of verification tasks. The infected models behave normally on benign samples, while attacker-specified unenrolled triggers will successfully pass the verification even if the attacker has no information about the enrolled speaker. We also demonstrate that existing back-door attacks cannot be directly adopted in attacking speaker verification. Our approach not only provides a new perspective for designing novel attacks, but also serves as a strong baseline for improving the robustness of verification methods. The code for reproducing main results is available at https://github.com/zhaitongqing233/Backdoor-attack-against-speaker-verification.

Cyber-defense systems are being developed to automatically ingest Cyber Threat Intelligence (CTI) that contains semi-structured data and/or text to populate knowledge graphs. A potential risk is that fake CTI can be generated and spread through Open-Source Intelligence (OSINT) communities or on the Web to effect a data poisoning attack on these systems. Adversaries can use fake CTI examples as training input to subvert cyber defense systems, forcing their models to learn incorrect inputs to serve the attackers' malicious needs. In this paper, we show how to automatically generate fake CTI text descriptions using transformers. Given an initial prompt sentence, a public language model like GPT-2 with fine-tuning can generate plausible CTI text that can mislead cyber-defense systems. We use the generated fake CTI text to perform a data poisoning attack on a Cybersecurity Knowledge Graph (CKG) and a cybersecurity corpus. The attack introduced adverse impacts such as returning incorrect reasoning outputs, representation poisoning, and corruption of other dependent AI-based cyber defense systems. We evaluate with traditional approaches and conduct a human evaluation study with cyber-security professionals and threat hunters. Based on the study, professional threat hunters were equally likely to consider our fake generated CTI and authentic CTI as true.

In machine learning Trojan attacks, an adversary trains a corrupted model that obtains good performance on normal data but behaves maliciously on data samples with certain trigger patterns. Several approaches have been proposed to detect such attacks, but they make undesirable assumptions about the attack strategies or require direct access to the trained models, which restricts their utility in practice.This paper addresses these challenges by introducing a Meta Neural Trojan Detection (MNTD) pipeline that does not make assumptions on the attack strategies and only needs black-box access to models. The strategy is to train a meta-classifier that predicts whether a given target model is Trojaned. To train the meta-model without knowledge of the attack strategy, we introduce a technique called jumbo learning that samples a set of Trojaned models following a general distribution. We then dynamically optimize a query set together with the meta-classifier to distinguish between Trojaned and benign models.We evaluate MNTD with experiments on vision, speech, tabular data and natural language text datasets, and against different Trojan attacks such as data poisoning attack, model manipulation attack, and latent attack. We show that MNTD achieves 97% detection AUC score and significantly outperforms existing detection approaches. In addition, MNTD generalizes well and achieves high detection performance against unforeseen attacks. We also propose a robust MNTD pipeline which achieves around 90% detection AUC even when the attacker aims to evade the detection with full knowledge of the system.

One of the most concerning threats for modern AI systems is data poisoning, where the attacker injects maliciously crafted training data to corrupt the system's behavior at test time. Availability poisoning is a particularly worrisome subset of poisoning attacks where the attacker aims to cause a Denial-of-Service (DoS) attack. However, the state-of-the-art algorithms are computationally expensive because they try to solve a complex bi-level optimization problem (the ``hammer''). We observed that in particular conditions, namely, where the target model is linear (the ``nut''), the usage of computationally costly procedures can be avoided. We propose a counter-intuitive but efficient heuristic that allows contaminating the training set such that the target system's performance is highly compromised. We further suggest a re-parameterization trick to decrease the number of variables to be optimized. Finally, we demonstrate that, under the considered settings, our framework achieves comparable, or even better, performances in terms of the attacker's objective while being significantly more computationally efficient.

Due to respectively limited training data, different entities addressing the same vision task based on certain sensitive images may not train a robust deep network. This paper introduces a new vision task where various entities share task-specific image data to enlarge each other's training data volume without visually disclosing sensitive contents (e.g. illegal images). Then, we present a new structure-based training regime to enable different entities learn task-specific and reconstruction-proof image representations for image data sharing. Specifically, each entity learns a private Deep Poisoning Module (DPM) and insert it to a pre-trained deep network, which is designed to perform the specific vision task. The DPM deliberately poisons convolutional image features to prevent image reconstructions, while ensuring that the altered image data is functionally equivalent to the non-poisoned data for the specific vision task. Given this equivalence, the poisoned features shared from one entity could be used by another entity for further model refinement. Experimental results on image classification prove the efficacy of the proposed method.

Federated learning (FL) is the rapidly developing machine learning technique that is used to perform collaborative model training over decentralized datasets. FL enables privacy-preserving model development whereby the datasets are scattered over a large set of data producers (i.e., devices and/or systems). These data producers train the learning models, encapsulate the model updates with differential privacy techniques, and share them to centralized systems for global aggregation. However, these centralized models are always prone to adversarial attacks (such as data-poisoning and model poisoning attacks) due to a large number of data producers. Hence, FL methods need to ensure fairness and high-quality model availability across all the participants in the underlying AI systems. In this paper, we propose a novel FL framework, called FairFed, to meet fairness and high-quality data requirements. The FairFed provides a fairness mechanism to detect adversaries across the devices and datasets in the FL network and reject their model updates. We use a Python-simulated FL framework to enable large-scale training over MNIST dataset. We simulate a cross-device model training settings to detect adversaries in the training network. We used TensorFlow Federated and Python to implement the fairness protocol, the deep neural network, and the outlier detection algorithm. We thoroughly test the proposed FairFed framework with random and uniform data distributions across the training network and compare our initial results with the baseline fairness scheme. Our proposed work shows promising results in terms of model accuracy and loss.

In today's technological climate, users require fast automation and digitization of results for large amounts of data at record speeds. Especially in the field of medicine, where each patient is often asked to undergo many different examinations within one diagnosis or treatment. Each examination can help in the diagnosis or prediction of further disease progression. Furthermore, all produced data from these examinations must be stored somewhere and available to various medical practitioners for analysis who may be in geographically diverse locations. The current medical climate leans towards remote patient monitoring and AI-assisted diagnosis. To make this possible, medical data should ideally be secured and made accessible to many medical practitioners, which makes them prone to malicious entities. Medical information has inherent value to malicious entities due to its privacy-sensitive nature in a variety of ways. Furthermore, if access to data is distributively made available to AI algorithms (particularly neural networks) for further analysis/diagnosis, the danger to the data may increase (e.g., model poisoning with fake data introduction). In this paper, we propose a federated learning approach that uses decentralized learning with blockchain-based security and a proposition that accompanies that training intelligent systems using distributed and locally-stored data for the use of all patients. Our work in progress hopes to contribute to the latest trend of the Internet of Medical Things security and privacy.

In recent years, Artificial Intelligence has disruptively changed information technology and software engineering with a proliferation of technologies and applications based-on it. However, recent researches show that AI models in general and the most greatest invention since sliced bread - Deep Learning models in particular, are vulnerable to being hacked and can be misused for bad purposes. In this paper, we carry out a brief review of data poisoning attack - one of the two recently dangerous emerging attacks - and the state-of-the-art defense methods for this problem. Finally, we discuss current challenges and future developments.

It has been recognized that artificial intelligence (AI) will play an important role in future societies. AI has already been incorporated in many industries to improve business processes and automation. Although the aviation industry has successfully implemented flight management systems or autopilot to automate flight operations, it is expected that full embracement of AI remains a challenge. Given the rigorous validation process and the requirements for the highest level of safety standards and risk management, AI needs to prove itself being safe to operate. This paper addresses the safety issues of AI deployment in an aviation network compatible with the Future Communication Infrastructure that utilizes heterogeneous wireless access technologies for communications between the aircraft and the ground networks. It further considers the exploitation of software defined networking (SDN) technologies in the ground network while the adoption of SDN in the airborne network can be optional. Due to the nature of centralized management in SDN-based network, the SDN controller can become a single point of failure or a target for cyber attacks. To countermeasure such attacks, an intrusion detection system utilises AI techniques, more specifically deep neural network (DNN), is considered. However, an adversary can target the AI-based intrusion detection system. This paper examines the impact of AI security attacks on the performance of the DNN algorithm. Poisoning attacks targeting the DSL-KDD datasets which were used to train the DNN algorithm were launched at the intrusion detection system. Results showed that the performance of the DNN algorithm has been significantly degraded in terms of the mean square error, accuracy rate, precision rate and the recall rate.

This paper discusses formulations and algorithms which allow a number of agents to collectively solve problems involving both (non-convex) minimization and (concave) maximization operations. These problems have a number of interesting applications in information processing and machine learning, and in particular can be used to model an adversary learning problem called network data poisoning. We develop a number of algorithms to efficiently solve these non-convex min-max optimization problems, by combining techniques such as gradient tracking in the decentralized optimization literature and gradient descent-ascent schemes in the min-max optimization literature. Also, we establish convergence to a first order stationary point under certain conditions. Finally, we perform experiments to demonstrate that the proposed algorithms are effective in the data poisoning attack.

Recently, deep neural network technology has been developed and used in various fields. The image recognition model can be used for automatic safety checks at the electric factory. However, as the deep neural network develops, the importance of security increases. A poisoning attack is one of security problems. It is an attack that breaks down by entering malicious data into the training data set of the model. This paper generates adversarial data that modulates feature values to different targets by manipulating less RGB values. Then, poisoning attacks in one of the image recognition models, the show and tell model. Then use autoencoder to defend adversarial data.

Few developments seem as poised to alter the characteristics of security in the digital age as the advent of artificial intelligence (AI) technologies. For national defense establishments, the emergence of AI techniques is particularly worrisome, not least because prototype applications already exist. Cyber attacks augmented by AI portend the tailored manipulation of human vectors within the attack surface of important societal systems at great scale, as well as opportunities for calamity resulting from the secondment of technical skill from the hacker to the algorithm. Arguably most important, however, is the fact that AI-enabled cyber campaigns contain great potential for operational obfuscation and strategic misdirection. At the operational level, techniques for piggybacking onto routine activities and for adaptive evasion of security protocols add uncertainty, complicating the defensive mission particularly where adversarial learning tools are employed in offense. Strategically, AI-enabled cyber operations offer distinct attempts to persistently shape the spectrum of cyber contention may be able to pursue conflict outcomes beyond the expected scope of adversary operation. On the other, AI-augmented cyber defenses incorporated into national defense postures are likely to be vulnerable to "poisoning" attacks that predict, manipulate and subvert the functionality of defensive algorithms. This article takes on two primary tasks. First, it considers and categorizes the primary ways in which AI technologies are likely to augment offensive cyber operations, including the shape of cyber activities designed to target AI systems. Then, it frames a discussion of implications for deterrence in cyberspace by referring to the policy of persistent engagement, agreed competition and forward defense promulgated in 2018 by the United States. Here, it is argued that the centrality of cyberspace to the deployment and operation of soon-to-be-ubiquitous AI systems implies new motivations for operation within the domain, complicating numerous assumptions that underlie current approaches. In particular, AI cyber operations pose unique measurement issues for the policy regime.

Machine-learning solutions are successfully adopted in multiple contexts but the application of these techniques to the cyber security domain is complex and still immature. Among the many open issues that affect security systems based on machine learning, we concentrate on adversarial attacks that aim to affect the detection and prediction capabilities of machine-learning models. We consider realistic types of poisoning and evasion attacks targeting security solutions devoted to malware, spam and network intrusion detection. We explore the possible damages that an attacker can cause to a cyber detector and present some existing and original defensive techniques in the context of intrusion detection systems. This paper contains several performance evaluations that are based on extensive experiments using large traffic datasets. The results highlight that modern adversarial attacks are highly effective against machine-learning classifiers for cyber detection, and that existing solutions require improvements in several directions. The paper paves the way for more robust machine-learning-based techniques that can be integrated into cyber security platforms.

Most of the data manipulation attacks on deep neural networks (DNNs) during the training stage introduce a perceptible noise that can be catered by preprocessing during inference, or can be identified during the validation phase. There-fore, data poisoning attacks during inference (e.g., adversarial attacks) are becoming more popular. However, many of them do not consider the imperceptibility factor in their optimization algorithms, and can be detected by correlation and structural similarity analysis, or noticeable (e.g., by humans) in multi-level security system. Moreover, majority of the inference attack rely on some knowledge about the training dataset. In this paper, we propose a novel methodology which automatically generates imperceptible attack images by using the back-propagation algorithm on pre-trained DNNs, without requiring any information about the training dataset (i.e., completely training data-unaware). We present a case study on traffic sign detection using the VGGNet trained on the German Traffic Sign Recognition Benchmarks dataset in an autonomous driving use case. Our results demonstrate that the generated attack images successfully perform misclassification while remaining imperceptible in both “subjective” and “objective” quality tests.

Deep neural networks are widely used in many walks of life. Techniques such as transfer learning enable neural networks pre-trained on certain tasks to be retrained for a new duty, often with much less data. Users have access to both pre-trained model parameters and model definitions along with testing data but have either limited access to training data or just a subset of it. This is risky for system-critical applications, where adversarial information can be maliciously included during the training phase to attack the system. Determining the existence and level of attack in a model is challenging. In this paper, we present evidence on how adversarially attacking training data increases the boundary of model parameters using as an example of a CNN model and the MNIST data set as a test. This expansion is due to new characteristics of the poisonous data that are added to the training data. Approaching the problem from the feature space learned by the network provides a relation between them and the possible parameters taken by the model on the training phase. An algorithm is proposed to determine if a given network was attacked in the training by comparing the boundaries of parameters distribution on intermediate layers of the model estimated by using the Maximum Entropy Principle and the Variational inference approach.

Federated learning is a novel distributed learning framework, where the deep learning model is trained in a collaborative manner among thousands of participants. The shares between server and participants are only model parameters, which prevent the server from direct access to the private training data. However, we notice that the federated learning architecture is vulnerable to an active attack from insider participants, called poisoning attack, where the attacker can act as a benign participant in federated learning to upload the poisoned update to the server so that he can easily affect the performance of the global model. In this work, we study and evaluate a poisoning attack in federated learning system based on generative adversarial nets (GAN). That is, an attacker first acts as a benign participant and stealthily trains a GAN to mimic prototypical samples of the other participants' training set which does not belong to the attacker. Then these generated samples will be fully controlled by the attacker to generate the poisoning updates, and the global model will be compromised by the attacker with uploading the scaled poisoning updates to the server. In our evaluation, we show that the attacker in our construction can successfully generate samples of other benign participants using GAN and the global model performs more than 80% accuracy on both poisoning tasks and main tasks.

Short-term load forecasting systems for power grids have demonstrated high accuracy and have been widely employed for commercial use. However, classic load forecasting systems, which are based on statistical methods, are subject to vulnerability from training data poisoning. In this paper, we demonstrate a data poisoning strategy that effectively corrupts the forecasting model even in the presence of outlier detection. To the best of our knowledge, poisoning attack on short-term load forecasting with outlier detection has not been studied in previous works. Our method applies to several forecasting models, including the most widely-adapted and best-performing ones, such as multiple linear regression (MLR) and neural network (NN) models. Starting with the MLR model, we develop a novel closed-form solution to quickly estimate the new MLR model after a round of data poisoning without retraining. We then employ line search and simulated annealing to find the poisoning attack solution. Furthermore, we use the MLR attacking solution to generate a numerical solution for other models, such as NN. The effectiveness of our algorithm has been tested on the Global Energy Forecasting Competition (GEFCom2012) data set with the presence of outlier detection.