Biblio

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.

Vehicular Ad hoc Networks commonly abbreviated as VANETs, are an important component of MANET. VANET refers to the group of vehicles that are interlinked to one another through wireless network. Along with technology, comes the threats. Like other wireless networks, VANETs also are vulnerable to various security threats. Security in VANETs is a major issue that attracted many researchers and academicians. One small security breach can cause a big damage in case of VANETs as in this case human lives are involved. Intrusion Detection Systems (IDS) are employed in VANETs in order to detect and identify any malicious activity in the network. The IDS works by analysing the network and detecting any intrusions tried or made in the network so that proper steps could be taken timely to prevent damage from such activities. This paper reviews Intrusion Detection systems, classification of IDS based on various factors and then the architecture of IDS. We then reviewed some of the recent and important intrusion detection research works and then compared them with one another.

As AI technology advances and its use increases, efforts to incorporate machine learning for malware detection are increasing. However, for malware learning, a standardized data set is required. Because malware is unstructured data, it cannot be directly learned. In order to solve this problem, many studies have attempted to convert unstructured data into structured data. In this study, the features and limitations of each were analyzed by investigating and analyzing the method of converting unstructured data proposed in each study into structured data. As a result, most of the data conversion techniques suggest conversion mechanisms, but the scope of each technique has not been determined. The resulting data set is not suitable for use as training data because it has infinite properties.

Face recognition has been used widely for personal authentication. However, there is a problem that it is vulnerable to a presentation attack in which a counterfeit such as a photo is presented to a camera to impersonate another person. Although various presentation attack detection methods have been proposed, these methods have not been able to sufficiently cope with the diversity of the heterogeneous environments including presentation attack instruments (PAIs) and lighting conditions. In this paper, we propose Learning by Environment Clusters (LEC) which divides training data into some clusters of similar photographic environments and trains bona-fide and attack classification models for each cluster. Experimental results using Replay-Attack, OULU-NPU, and CelebA-Spoof show the EER of the conventional method which trains one classification model from all data was 20.0%, but LEC can achieve 13.8% EER when using binarized statistical image features (BSIFs) and support vector machine used as the classification method.

Sound plays a key role in human life and therefore sound recognition system has a great future ahead. Sound classification and identification system has many applications such as system for personal security, critical surveillance, etc. The main aim of this paper is to detect and classify the security sound event using the surveillance camera systems with integrated microphone based on the generated spectrograms of the sounds. This will enable to track security events in cases of emergencies. The goal is to propose a security system to accurately detect sound events and make a better security sound event detection system. We propose to use a convolutional neural network (CNN) to design the security sound detection system to detect a security event with minimal sound. We used the spectrogram images to train the CNN. The neural network was trained using different security sounds data which was then used to detect security sound events during testing phase. We used two datasets for our experiment training and testing datasets. Both the datasets contain 3 different sound events (glass break, gun shots and smoke alarms) to train and test the model, respectively. The proposed system yields the good accuracy for the sound event detection even with minimum available sound data. The designed system achieved accuracy was 92% and 90% using CNN on training dataset and testing dataset. We conclude that the proposed sound classification framework which using the spectrogram images of sounds can be used efficiently to develop the sound classification and recognition systems.

Federated learning is a privacy-preserving learning system where participants locally update a shared model with their own training data. Despite the advantage that training data are not sent to a server, there is still a risk that a state-of-the-art model inversion attack, which may be conducted by the server, infers training data from the models updated by the participants, referred to as individual models. A solution to prevent such attacks is differential privacy, where each participant adds noise to the individual model before sending it to the server. Differential privacy, however, sacrifices the quality of the shared model in compensation for the fact that participants' training data are not leaked. This paper proposes a federated learning system that is resistant to model inversion attacks without sacrificing the quality of the shared model. The core idea is that each participant divides the individual model into model fragments, shuffles, and aggregates them to prevent adversaries from inferring training data. The other benefit of the proposed system is that the resulting shared model is identical to the shared model generated with the naive federated learning.

An adversarial attack is an algorithm that perturbs the input of a machine learning model in an intelligent way in order to change the output of the model. An important property of adversarial attacks is transferability. According to this property, it is possible to generate adversarial perturbations on one model and apply it the input to fool the output of a different model. Our work focuses on studying the transferability of adversarial attacks in sound event classification. We are able to demonstrate differences in transferability properties from those observed in computer vision. We show that dataset normalization techniques such as z-score normalization does not affect the transferability of adversarial attacks and we show that techniques such as knowledge distillation do not increase the transferability of attacks.

Classification between machine-translated text and original text are often tokenized on vocabulary of the corpi. With N-grams larger than uni-gram, one can create a model that estimates a decision boundary based on word frequency probability distribution; however, this approach is exponentially expensive because of high dimensionality and sparsity. Instead, we let samples of the corpi be represented by part-of-speech tagging which is significantly less vocabulary. With less trigram permutations, we can create a model with its tri-gram frequency probability distribution. In this paper, we explore less conventional ways of approaching techniques for handling documents, dictionaries, and the likes.

Most machine learning-based malware detection systems use various supervised learning methods to classify different instances of software as benign or malicious. This approach provides no information regarding the behavioral characteristics of malware. It also requires a large amount of training data and is prone to labeling difficulties and can reduce accuracy due to redundant training data. Therefore, we propose a malware detection method based on deep learning, which uses malware images and a set of autoencoders to detect malware. The method is to design an autoencoder to learn the functional characteristics of malware, and then to observe the reconstruction error of autoencoder to realize the classification and detection of malware and benign software. The proposed approach achieves 93% accuracy and comparatively better F1-score values while detecting malware and needs little training data when compared with traditional malware detection systems.

Side-channel attacks occur on smartphone keystrokes, where the input can be intercepted by a tapping sound. Ilia et al. reported that keystrokes can be predicted with 61% accuracy from tapping sounds listened to by the built-in microphone of a legitimate user's device. Li et al. reported that by emitting sonar sounds from an attacker smartphone's built-in speaker and analyzing the reflected waves from a legitimate user's finger at the time of tap input, keystrokes can be estimated with 90% accuracy. However, the method proposed by Ilia et al. requires prior penetration of the target smartphone and the attack scenario lacks plausibility; if the attacker's smartphone can be penetrated, the keylogger can directly acquire the keystrokes of a legitimate user. In addition, the method proposed by Li et al. is a side-channel attack in which the attacker actively interferes with the terminals of legitimate users and can be described as an active attack scenario. Herein, we analyze the extent to which a user's keystrokes are leaked to the attacker in a passive attack scenario, where the attacker wiretaps the sounds of the legitimate user's keystrokes using an external microphone. First, we limited the keystrokes to the personal identification number input. Subsequently, mel-frequency cepstrum coefficients of tapping sound data were represented as image data. Consequently, we found that the input is discriminated with high accuracy using a convolutional neural network to estimate the key input.

The completion and the correction of measurement data are the foundation of the ubiquitous power internet of things construction. However, data missing may occur during the data transporting process. Therefore, a model of missing load situation reconstruction based on the generative adversarial networks is proposed in this paper to overcome the disadvantage of depending on data of other relevant factors in conventional methods. Through the unsupervised training, the proposed model can automatically learn the complex features of loads that are difficult to model explicitly to fill the incomplete load data without using other relevant data. Meanwhile, a method of online correction is put forward to improve the robustness of the reconstruction model in different scenarios. The proposed method is fully data-driven and contains no explicit modeling process. The test results indicate that the proposed algorithm is well-matched for the various scenarios, including the discontinuous missing load reconstruction and the continuous missing load reconstruction even massive data missing. Specifically, the reconstruction error rate of the proposed algorithm is within 4% under the absence of 50% load data.

State estimation is of considerable significance for the power system operation and control. However, well-designed false data injection attacks can utilize blind spots in conventional residual-based bad data detection methods to manipulate measurements in a coordinated manner and thus affect the secure operation and economic dispatch of grids. In this paper, we propose a detection approach based on an autoencoder neural network. By training the network on the dependencies intrinsic in `normal' operation data, it effectively overcomes the challenge of unbalanced training data that is inherent in power system attack detection. To evaluate the detection performance of the proposed mechanism, we conduct a series of experiments on the IEEE 118-bus power system. The experiments demonstrate that the proposed autoencoder detector displays robust detection performance under a variety of attack scenarios.

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.

Online detection of anomalies is crucial to enhancing the reliability and resiliency of power systems. We propose a novel data-driven online event detection algorithm with synchrophasor data using graph signal processing. In addition to being extremely scalable, our proposed algorithm can accurately capture and leverage the spatio-temporal correlations of the streaming PMU data. This paper also develops a general technique to decouple spatial and temporal correlations in multiple time series. Finally, we develop a unique framework to construct a weighted adjacency matrix and graph Laplacian for product graph. Case studies with real-world, large-scale synchrophasor data demonstrate the scalability and accuracy of our proposed event detection algorithm. Compared to the state-of-the-art benchmark, the proposed method not only achieves higher detection accuracy but also yields higher computational efficiency.

In this paper, we assess the vulnerability of network anomaly detection systems that use machine learning methods. Although the performance of these network anomaly detection systems is high in comparison to that of existing methods without machine learning methods, the use of machine learning methods for detecting vulnerabilities is a growing concern among researchers of image processing. If the vulnerabilities of machine learning used in the network anomaly detection method are exploited by attackers, large security threats are likely to emerge in the near future. Therefore, in this paper we clarify how vulnerability detection of machine learning network anomaly detection methods affects their performance.

Recent advances in generative adversarial networks have made detecting fake videos a challenging task. In this paper, we propose the application of the state-of-the-art attribution based confidence (ABC) metric for detecting deepfake videos. The ABC metric does not require access to the training data or training the calibration model on the validation data. The ABC metric can be used to draw inferences even when only the trained model is available. Here, we utilize the ABC metric to characterize whether a video is original or fake. The deep learning model is trained only on original videos. The ABC metric uses the trained model to generate confidence values. For, original videos, the confidence values are greater than 0.94.

In the last couple of years, the move to cyberspace provides a fertile environment for ransomware criminals like ever before. Notably, since the introduction of WannaCry, numerous ransomware detection solution has been proposed. However, the ransomware incidence report shows that most organizations impacted by ransomware are running state of the art ransomware detection tools. Hence, an alternative solution is an urgent requirement as the existing detection models are not sufficient to spot emerging ransomware treat. With this motivation, our work proposes "DeepGuard," a novel concept of modeling user behavior for ransomware detection. The main idea is to log the file-interaction pattern of typical user activity and pass it through deep generative autoencoder architecture to recreate the input. With sufficient training data, the model can learn how to reconstruct typical user activity (or input) with minimal reconstruction error. Hence, by applying the three-sigma limit rule on the model's output, DeepGuard can distinguish the ransomware activity from the user activity. The experiment result shows that DeepGuard effectively detects a variant class of ransomware with minimal false-positive rates. Overall, modeling the attack detection with user-behavior permits the proposed strategy to have deep visibility of various ransomware families.

In recent years, artificial intelligence has been widely used in the field of network security, which has significantly improved the effect of network security analysis and detection. However, because the power industrial control system is faced with the problem of shortage of attack data, the direct deployment of the network intrusion detection system based on artificial intelligence is faced with the problems of lack of data, low precision, and high false alarm rate. To solve this problem, we propose an anomaly traffic detection method based on cross-domain knowledge transferring. By using the TrAdaBoost algorithm, we achieve a lower error rate than using LSTM alone.

Over the last few years, there has been an increasing number of studies about facial emotion recognition because of the importance and the impact that it has in the interaction of humans with computers. With the growing number of challenging datasets, the application of deep learning techniques have all become necessary. In this paper, we study the challenges of Emotion Recognition Datasets and we also try different parameters and architectures of the Conventional Neural Networks (CNNs) in order to detect the seven emotions in human faces, such as: anger, fear, disgust, contempt, happiness, sadness and surprise. We have chosen iCV MEFED (Multi-Emotion Facial Expression Dataset) as the main dataset for our study, which is relatively new, interesting and very challenging.

This paper proposes a deep learning-based white-hat worm launcher in Botnet Defense System (BDS). BDS uses white-hat botnets to defend an IoT system against malicious botnets. White-hat worm launcher literally launches white-hat worms to create white-hat botnets according to the strategy decided by BDS. The proposed launcher learns with deep learning where is the white-hat worms' right place to successfully drive out malicious botnets. Given a system situation invaded by malicious botnets, it predicts a worms' placement by the learning result and launches them. We confirmed the effect of the proposed launcher through simulating evaluation.

In Machine Learning, White Box Adversarial Attacks rely on knowing underlying knowledge about the model attributes. This works focuses on discovering to distrinct pieces of model information: the underlying architecture and primary training dataset. With the process in this paper, a structured set of input probes and the output of the model become the training data for a deep classifier. Two subdomains in Machine Learning are explored - image based classifiers and text transformers with GPT-2. With image classification, the focus is on exploring commonly deployed architectures and datasets available in popular public libraries. Using a single transformer architecture with multiple levels of parameters, text generation is explored by fine tuning off different datasets. Each dataset explored in image and text are distinguishable from one another. Diversity in text transformer outputs implies further research is needed to successfully classify architecture attribution in text domain.

With the advent of networking technologies and increasing network attacks, Intrusion Detection systems are apparently needed to stop attacks and malicious activities. Various frameworks and techniques have been developed to solve the problem of intrusion detection, still there is need for new frameworks as per the challenging scenario of enormous scale in data size and nature of attacks. Current IDS systems pose challenges on the throughput to work with high speed networks. In this paper we address the issue of high computational overhead of anomaly based IDS and propose the solution using discretization as a data preprocessing step which can drastically reduce the computation overhead. We propose method to provide near real time detection of attacks using only basic flow level features that can easily be extracted from network packets.

Reinforcement learning (RL) algorithms have been demonstrated to be very attractive tools to train agents to achieve sequential tasks. However, these algorithms require too many training data to converge to be efficiently applied to physical robots. By using a human teacher, the learning process can be made faster and more robust, but the overall performance heavily depends on the quality and availability of teacher demonstrations or instructions. In particular, when these teaching signals are inadequate, the agent may fail to learn an optimal policy. In this paper, we introduce a trust-based interactive task learning approach. We propose an RL architecture able to learn both from environment rewards and from various sparse teaching signals provided by non-expert teachers, using an actor-critic agent, a human model and a trust model. We evaluate the performance of this architecture on 4 different setups using a maze environment with different simulated teachers and show that the benefits of the trust model.

Detection of multi-stage threats such as Advanced Persistent Threats (APT) is extremely challenging due to their deceptive approaches. Sequential events of threats might look benign when performed individually or from different addresses. We propose a new unsupervised framework to identify patterns and correlations of malicious behaviours by analysing heterogeneous log-files. The framework consists of two main phases of data analysis to extract inner-behaviours of log-files and then the patterns of those behaviours over analysed files. To evaluate the framework we have produced a (publicly available) labelled version of the SotM43 dataset. Our results demonstrate that the framework can (i) efficiently cluster inner-behaviours of log-files with high accuracy and (ii) extract patterns of malicious behaviour and correlations between those patterns from real-world data.