Visible to the public Biblio

Found 1564 results

Filters: First Letter Of Last Name is D  [Clear All Filters]
Dykstra, J..  2015.  Essential Cybersecurity Science: Build, Test, and Evaluate Secure Systems. :190.

If you’re involved in cybersecurity as a software developer, forensic investigator, or network administrator, this practical guide shows you how to apply the scientific method when assessing techniques for protecting your information systems. You’ll learn how to conduct scientific experiments on everyday tools and procedures, whether you’re evaluating corporate security systems, testing your own security product, or looking for bugs in a mobile game.

Once author Josiah Dykstra gets you up to speed on the scientific method, he helps you focus on standalone, domain-specific topics, such as cryptography, malware analysis, and system security engineering. The latter chapters include practical case studies that demonstrate how to use available tools to conduct domain-specific scientific experiments.

  • Learn the steps necessary to conduct scientific experiments in cybersecurity
  • Explore fuzzing to test how your software handles various inputs
  • Measure the performance of the Snort intrusion detection system
  • Locate malicious “needles in a haystack” in your network and IT environment
  • Evaluate cryptography design and application in IoT products
  • Conduct an experiment to identify relationships between similar malware binaries
  • Understand system-level security requirements for enterprise networks and web services
Dylan Wang, Melody Moh, Teng-Sheng Moh.  2020.  Using Deep Learning to Solve Google reCAPTCHA v2’s Image Challenges.

The most popular CAPTCHA service in use today is Google reCAPTCHA v2, whose main offering is an image-based CAPTCHA challenge. This paper looks into the security measures used in reCAPTCHA v2's image challenges and proposes a deep learning-based solution that can be used to automatically solve them. The proposed method is tested with both a custom object- detection deep learning model as well as Google's own Cloud Vision API, in conjunction with human mimicking mouse movements to bypass the challenges. The paper also suggests some potential defense measures to increase overall security and other additional attack directions for reCAPTCHA v2.

Book Chapter
Conference Paper
Doerr, Carola, Lengler, Johannes.  2016.  The (1+1) Elitist Black-Box Complexity of LeadingOnes. Proceedings of the Genetic and Evolutionary Computation Conference 2016. :1131–1138.

One important goal of black-box complexity theory is the development of complexity models allowing to derive meaningful lower bounds for whole classes of randomized search heuristics. Complementing classical runtime analysis, black-box models help us understand how algorithmic choices such as the population size, the variation operators, or the selection rules influence the optimization time. One example for such a result is the Ω(n log n) lower bound for unary unbiased algorithms on functions with a unique global optimum [Lehre/Witt, GECCO 2010], which tells us that higher arity operators or biased sampling strategies are needed when trying to beat this bound. In lack of analyzing techniques, almost no non-trivial bounds are known for other restricted models. Proving such bounds therefore remains to be one of the main challenges in black-box complexity theory. With this paper we contribute to our technical toolbox for lower bound computations by proposing a new type of information-theoretic argument. We regard the permutation- and bit-invariant version of LeadingOnes and prove that its (1+1) elitist black-box complexity is Ω(n2), a bound that is matched by (1+1)-type evolutionary algorithms. The (1+1) elitist complexity of LeadingOnes is thus considerably larger than its unrestricted one, which is known to be of order n log log n [Afshani et al., 2013].

Demjaha, A., Caulfield, T., Sasse, M. Angela, Pym, D..  2019.  2 Fast 2 Secure: A Case Study of Post-Breach Security Changes. 2019 IEEE European Symposium on Security and Privacy Workshops (EuroS PW). :192—201.
A security breach often makes companies react by changing their attitude and approach to security within the organization. This paper presents an in-depth case study of post-breach security changes made by a company and the consequences of those changes. We employ the principles of participatory action research and humble inquiry to conduct a long-term study with employee interviews while embedded in the organization's security division. Despite an extremely high level of financial investment in security, and consistent attention and involvement from the board, the interviews indicate a significant level of friction between employees and security. In the main themes that emerged from our data analysis, a number of factors shed light on the friction: fear of another breach leading to zero risk appetite, impossible security controls making non-compliance a norm, security theatre underminining the purpose of security policies, employees often trading-off security with productivity, and as such being treated as children in detention rather than employees trying to finish their paid jobs. This paper shows that post-breach security changes can be complex and sometimes risky due to emotions often being involved. Without an approach considerate of how humans and security interact, even with high financial investment, attempts to change an organization's security behaviour may be ineffective.
Ly, Son Thai, Do, Nhu-Tai, Lee, Guee-Sang, Kim, Soo-Hyung, Yang, Hyung-Jeong.  2019.  A 3d Face Modeling Approach for in-The-Wild Facial Expression Recognition on Image Datasets. 2019 IEEE International Conference on Image Processing (ICIP). :3492—3496.

This paper explores the benefits of 3D face modeling for in-the-wild facial expression recognition (FER). Since there is limited in-the-wild 3D FER dataset, we first construct 3D facial data from available 2D dataset using recent advances in 3D face reconstruction. The 3D facial geometry representation is then extracted by deep learning technique. In addition, we also take advantage of manipulating the 3D face, such as using 2D projected images of 3D face as additional input for FER. These features are then fused with that of 2D FER typical network. By doing so, despite using common approaches, we achieve a competent recognition accuracy on Real-World Affective Faces (RAF) database and Static Facial Expressions in the Wild (SFEW 2.0) compared with the state-of-the-art reports. To the best of our knowledge, this is the first time such a deep learning combination of 3D and 2D facial modalities is presented in the context of in-the-wild FER.

Dong, C., Liu, Y., Zhang, Y., Shi, P., Shao, X., Ma, C..  2018.  Abnormal Bus Data Detection of Intelligent and Connected Vehicle Based on Neural Network. 2018 IEEE International Conference on Computational Science and Engineering (CSE). :171–176.
In the paper, our research of abnormal bus data analysis of intelligent and connected vehicle aims to detect the abnormal data rapidly and accurately generated by the hackers who send malicious commands to attack vehicles through three patterns, including remote non-contact, short-range non-contact and contact. The research routine is as follows: Take the bus data of 10 different brands of intelligent and connected vehicles through the real vehicle experiments as the research foundation, set up the optimized neural network, collect 1000 sets of the normal bus data of 15 kinds of driving scenarios and the other 300 groups covering the abnormal bus data generated by attacking the three systems which are most common in the intelligent and connected vehicles as the training set. In the end after repeated amendments, with 0.5 seconds per detection, the intrusion detection system has been attained in which for the controlling system the abnormal bus data is detected at the accuracy rate of 96% and the normal data is detected at the accuracy rate of 90%, for the body system the abnormal one is 87% and the normal one is 80%, for the entertainment system the abnormal one is 80% and the normal one is 65%.
Donghoon Kim, Henry E. Schaffer, Mladen Vouk.  2015.  About PaaS Security. 3rd International IBM Cloud Academy Conference (ICACON 2015).
Duraisamy, Karthi, Lu, Hao, Pande, Partha Pratim, Kalyanaraman, Ananth.  2017.  Accelerating Graph Community Detection with Approximate Updates via an Energy-Efficient NoC. Proceedings of the 54th Annual Design Automation Conference 2017. :89:1–89:6.

Community detection is an advanced graph operation that is used to reveal tightly-knit groups of vertices (aka. communities) in real-world networks. Given the intractability of the problem, efficient heuristics are used in practice. Yet, even the best of these state-of-the-art heuristics can become computationally demanding over large inputs and can generate workloads that exhibit inherent irregularity in data movement on manycore platforms. In this paper, we posit that effective acceleration of the graph community detection operation can be achieved by reducing the cost of data movement through a combined innovation at both software and hardware levels. More specifically, we first propose an efficient software-level parallelization of community detection that uses approximate updates to cleverly exploit a diminishing returns property of the algorithm. Secondly, as a way to augment this innovation at the software layer, we design an efficient Wireless Network on Chip (WiNoC) architecture that is suited to handle the irregular on-chip data movements exhibited by the community detection algorithm under both unicast- and broadcast-heavy cache coherence protocols. Experimental results show that our resulting WiNoC-enabled manycore platform achieves on average 52% savings in execution time, without compromising on the quality of the outputs, when compared to a traditional manycore platform designed with a wireline mesh NoC and running community detection without employing approximate updates.

Bai, Xu, Jiang, Lei, Dai, Qiong, Yang, Jiajia, Tan, Jianlong.  2017.  Acceleration of RSA processes based on hybrid ARM-FPGA cluster. 2017 IEEE Symposium on Computers and Communications (ISCC). :682–688.

Cooperation of software and hardware with hybrid architectures, such as Xilinx Zynq SoC combining ARM CPU and FPGA fabric, is a high-performance and low-power platform for accelerating RSA Algorithm. This paper adopts the none-subtraction Montgomery algorithm and the Chinese Remainder Theorem (CRT) to implement high-speed RSA processors, and deploys a 48-node cluster infrastructure based on Zynq SoC to achieve extremely high scalability and throughput of RSA computing. In this design, we use the ARM to implement node-to-node communication with the Message Passing Interface (MPI) while use the FPGA to handle complex calculation. Finally, the experimental results show that the overall performance is linear with the number of nodes. And the cluster achieves 6× 9× speedup against a multi-core desktop (Intel i7-3770) and comparable performance to a many-core server (288-core). In addition, we gain up to 2.5× energy efficiency compared to these two traditional platforms.

Yi, Zhuo, Du, Xuehui, Liao, Ying, Lu, Xin.  2019.  An Access Authentication Algorithm Based on a Hierarchical Identity-Based Signature over Lattice for the Space-Ground Integrated Network. 2019 International Conference on Advanced Communication Technologies and Networking (CommNet). :1–9.

Access authentication is a key technology to identify the legitimacy of mobile users when accessing the space-ground integrated networks (SGIN). A hierarchical identity-based signature over lattice (L-HIBS) based mobile access authentication mechanism is proposed to settle the insufficiencies of existing access authentication methods in SGIN such as high computational complexity, large authentication delay and no-resistance to quantum attack. Firstly, the idea of hierarchical identity-based cryptography is introduced according to hierarchical distribution of nodes in SGIN, and a hierarchical access authentication architecture is built. Secondly, a new L-HIBS scheme is constructed based on the Small Integer Solution (SIS) problem to support the hierarchical identity-based cryptography. Thirdly, a mobile access authentication protocol that supports bidirectional authentication and shared session key exchange is designed with the aforementioned L-HIBS scheme. Results of theoretical analysis and simulation experiments suggest that the L-HIBS scheme possesses strong unforgeability of selecting identity and adaptive selection messages under the standard security model, and the authentication protocol has smaller computational overhead and shorter private keys and shorter signature compared to given baseline protocols.

Vanickis, R., Jacob, P., Dehghanzadeh, S., Lee, B..  2018.  Access Control Policy Enforcement for Zero-Trust-Networking. 2018 29th Irish Signals and Systems Conference (ISSC). :1-6.

The evolution of the enterprise computing landscape towards emerging trends such as fog/edge computing and the Industrial Internet of Things (IIoT) are leading to a change of approach to securing computer networks to deal with challenges such as mobility, virtualized infrastructures, dynamic and heterogeneous user contexts and transaction-based interactions. The uncertainty introduced by such dynamicity introduces greater uncertainty into the access control process and motivates the need for risk-based access control decision making. Thus, the traditional perimeter-based security paradigm is increasingly being abandoned in favour of a so called "zero trust networking" (ZTN). In ZTN networks are partitioned into zones with different levels of trust required to access the zone resources depending on the assets protected by the zone. All accesses to sensitive information is subject to rigorous access control based on user and device profile and context. In this paper we outline a policy enforcement framework to address many of open challenges for risk-based access control for ZTN. We specify the design of required policy languages including a generic firewall policy language to express firewall rules. We design a mechanism to map these rules to specific firewall syntax and to install the rules on the firewall. We show the viability of our design with a small proof-of-concept.

Datta, A., Kar, S., Sinopoli, B., Weerakkody, S..  2016.  Accountability in cyber-physical systems. 2016 Science of Security for Cyber-Physical Systems Workshop (SOSCYPS). :1–3.

Our position is that a key component of securing cyber-physical systems (CPS) is to develop a theory of accountability that encompasses both control and computing systems. We envision that a unified theory of accountability in CPS can be built on a foundation of causal information flow analysis. This theory will support design and analysis of mechanisms at various stages of the accountability regime: attack detection, responsibility-assignment (e.g., attack identification or localization), and corrective measures (e.g., via resilient control) As an initial step in this direction, we summarize our results on attack detection in control systems. We use the Kullback-Liebler (KL) divergence as a causal information flow measure. We then recover, using information flow analyses, a set of existing results in the literature that were previously proved using different techniques. These results cover passive detection, stealthy attack characterization, and active detection. This research direction is related to recent work on accountability in computational systems [1], [2], [3], [4]. We envision that by casting accountability theories in computing and control systems in terms of causal information flow, we can provide a common foundation to develop a theory for CPS that compose elements from both domains.

Baldimtsi, F., Camenisch, J., Dubovitskaya, M., Lysyanskaya, A., Reyzin, L., Samelin, K., Yakoubov, S..  2017.  Accumulators with Applications to Anonymity-Preserving Revocation. 2017 IEEE European Symposium on Security and Privacy (EuroS P). :301–315.

Membership revocation is essential for cryptographic applications, from traditional PKIs to group signatures and anonymous credentials. Of the various solutions for the revocation problem that have been explored, dynamic accumulators are one of the most promising. We propose Braavos, a new, RSA-based, dynamic accumulator. It has optimal communication complexity and, when combined with efficient zero-knowledge proofs, provides an ideal solution for anonymous revocation. For the construction of Braavos we use a modular approach: we show how to build an accumulator with better functionality and security from accumulators with fewer features and weaker security guarantees. We then describe an anonymous revocation component (ARC) that can be instantiated using any dynamic accumulator. ARC can be added to any anonymous system, such as anonymous credentials or group signatures, in order to equip it with a revocation functionality. Finally, we implement ARC with Braavos and plug it into Idemix, the leading implementation of anonymous credentials. This work resolves, for the first time, the problem of practical revocation for anonymous credential systems.

Darabseh, A., Namin, A. S..  2015.  On Accuracy of Classification-Based Keystroke Dynamics for Continuous User Authentication. 2015 International Conference on Cyberworlds (CW). :321–324.

The aim of this research is to advance the user active authentication using keystroke dynamics. Through this research, we assess the performance and influence of various keystroke features on keystroke dynamics authentication systems. In particular, we investigate the performance of keystroke features on a subset of most frequently used English words. The performance of four features such as i) key duration, ii) flight time latency, iii) diagraph time latency, and iv) word total time duration are analyzed. Two machine learning techniques are employed for assessing keystroke authentications. The selected classification methods are support vector machine (SVM), and k-nearest neighbor classifier (K-NN). The logged experimental data are captured for 28 users. The experimental results show that key duration time offers the best performance result among all four keystroke features, followed by word total time.

Almogbil, Atheer, Alghofaili, Abdullah, Deane, Chelsea, Leschke, Timothy, Almogbil, Atheer, Alghofaili, Abdullah.  2020.  The Accuracy of GPS-Enabled Fitbit Activities as Evidence: A Digital Forensics Study. 2020 7th IEEE International Conference on Cyber Security and Cloud Computing (CSCloud)/2020 6th IEEE International Conference on Edge Computing and Scalable Cloud (EdgeCom). :186—189.

Technology is advancing rapidly and with this advancement, it has become apparent that it is nearly impossible to not leave a digital trace when committing a crime. As evidenced by multiple cases handled by law enforcement, Fitbit data has proved to be useful when determining the validity of alibis and in piecing together the timeline of a crime scene. In our paper, experiments testing the accuracy and reliability of GPS-tracked activities logged by the Fitbit Alta tracker and Ionic smartwatch are conducted. Potential indicators of manipulated or altered GPS-tracked activities are identified to help guide digital forensic investigators when handling such Fitbit data as evidence.

Darabseh, A., Namin, A. Siami.  2015.  On Accuracy of Keystroke Authentications Based on Commonly Used English Words. 2015 International Conference of the Biometrics Special Interest Group (BIOSIG). :1–8.

The aim of this research is to advance the user active authentication using keystroke dynamics. Through this research, we assess the performance and influence of various keystroke features on keystroke dynamics authentication systems. In particular, we investigate the performance of keystroke features on a subset of most frequently used English words. The performance of four features such as i) key duration, ii) flight time latency, iii) digraph time latency, and iv) word total time duration are analyzed. Experiments are performed to measure the performance of each feature individually as well as the results from the different subsets of these features. Four machine learning techniques are employed for assessing keystroke authentications. The selected classification methods are two-class support vector machine (TC) SVM, one-class support vector machine (OC) SVM, k-nearest neighbor classifier (K-NN), and Naive Bayes classifier (NB). The logged experimental data are captured for 28 users. The experimental results show that key duration time offers the best performance result among all four keystroke features, followed by word total time. Furthermore, our results show that TC SVM and KNN perform the best among the four classifiers.

Copty, Fady, Danos, Matan, Edelstein, Orit, Eisner, Cindy, Murik, Dov, Zeltser, Benjamin.  2018.  Accurate Malware Detection by Extreme Abstraction. Proceedings of the 34th Annual Computer Security Applications Conference. :101–111.

Modern malware applies a rich arsenal of evasion techniques to render dynamic analysis ineffective. In turn, dynamic analysis tools take great pains to hide themselves from malware; typically this entails trying to be as faithful as possible to the behavior of a real run. We present a novel approach to malware analysis that turns this idea on its head, using an extreme abstraction of the operating system that intentionally strays from real behavior. The key insight is that the presence of malicious behavior is sufficient evidence of malicious intent, even if the path taken is not one that could occur during a real run of the sample. By exploring multiple paths in a system that only approximates the behavior of a real system, we can discover behavior that would often be hard to elicit otherwise. We aggregate features from multiple paths and use a funnel-like configuration of machine learning classifiers to achieve high accuracy without incurring too much of a performance penalty. We describe our system, TAMALES (The Abstract Malware Analysis LEarning System), in detail and present machine learning results using a 330K sample set showing an FPR (False Positive Rate) of 0.10% with a TPR (True Positive Rate) of 99.11%, demonstrating that extreme abstraction can be extraordinarily effective in providing data that allows a classifier to accurately detect malware.

Gordon, Kiel, Davis, Matthew, Birnbaum, Zachary, Dolgikh, Andrey.  2018.  ACE: Advanced CIP Evaluator. Proceedings of the 2018 Workshop on Cyber-Physical Systems Security and PrivaCy. :90-101.

Industrial control systems (ICS) are key enabling systems that drive the productivity and efficiency of omnipresent industries such as power, gas, water treatment, transportation, and manufacturing. These systems consist of interconnected components that communicate over industrial networks using industrial protocols such as the Common Industrial Protocol (CIP). CIP is one of the most commonly used network-based process control protocols, and utilizes an object-oriented communication structure for device to device interaction. Due to this object-oriented structure, CIP communication reveals detailed information about the devices, the communication patterns, and the system, providing an in-depth view of the system. The details from this in-depth system perspective can be utilized as part of a system cybersecurity or discovery approach. However, due to the variety of commands, corresponding parameters, and variable layer structure of the CIP network layer, processing this layer is a challenging task. This paper presents a tool, Advanced CIP Evaluator (ACE), which passively processes the CIP communication layer and automatically extracts device, communication, and system information from observed network traffic. ACE was tested and verified using a representative ICS power generation testbed. Since ACE operates passively, without generating any network traffic of its own, system operations are not disturbed. This novel tool provides ICS information, such as networked devices, communication patterns, and system operation, at a depth and breadth that is unique compared with other known tools.

Dong, Qiuxiang, Huang, Dijiang, Luo, Jim, Kang, Myong.  2018.  Achieving Fine-Grained Access Control with Discretionary User Revocation over Cloud Data. 2018 IEEE Conference on Communications and Network Security (CNS). :1—9.
Cloud storage solutions have gained momentum in recent years. However, cloud servers can not be fully trusted. Data access control have becomes one of the main impediments for further adoption. One appealing approach is to incorporate the access control into encrypted data, thus removing the need to trust the cloud servers. Among existing cryptographic solutions, Ciphertext Policy Attribute-Based Encryption (CP-ABE) is well suited for fine-grained data access control in cloud storage. As promising as it is, user revocation is a cumbersome problem that impedes its wide application. To address this issue, we design an access control system called DUR-CP-ABE, which implements identity-based User Revocation in a data owner Discretionary way. In short, the proposed solution provides the following salient features. First, user revocation enforcement is based on the discretion of the data owner, thus providing more flexibility. Second, no private key updates are needed when user revocation occurs. Third, the proposed scheme allows for group revocation of affiliated users in a batch operation. To the best of our knowledge, DUR-CP-ABE is the first CP-ABE solution to provide affiliation- based batch revocation functionality, which fits naturally into organizations' Identity and Access Management (IAM) structure. The analysis shows that the proposed access control system is provably secure and efficient in terms of computation, communi- cation and storage.
Fang, Y., Dickerson, S. J..  2017.  Achieving Swarm Intelligence with Spiking Neural Oscillators. 2017 IEEE International Conference on Rebooting Computing (ICRC). :1–4.

Mimicking the collaborative behavior of biological swarms, such as bird flocks and ant colonies, Swarm Intelligence algorithms provide efficient solutions for various optimization problems. On the other hand, a computational model of the human brain, spiking neural networks, has been showing great promise in recognition, inference, and learning, due to recent emergence of neuromorphic hardware for high-efficient and low-power computing. Through bridging these two distinct research fields, we propose a novel computing paradigm that implements the swarm intelligence with a population of coupled spiking neural oscillators in basic leaky integrate-and-fire (LIF) model. Our model behaves as a meta-heuristic searching conducted by multiple collaborative agents. In this design, the oscillating neurons serve as agents in the swarm, search for solutions in frequency coding and communicate with each other through spikes. The firing rate of each agent is adaptive to other agents with better solutions and the optimal solution is rendered as the swarm synchronization is reached. We apply the proposed method to the parameter optimization in several test objective functions and demonstrate its effectiveness and efficiency. Our new computing paradigm expands the computational power of coupled spiking neurons in the field of solving optimization problem and brings opportunities for the connection between individual intelligence and swarm intelligence.

Huang, Lulu, Matwin, Stan, de Carvalho, Eder J., Minghim, Rosane.  2017.  Active Learning with Visualization for Text Data. Proceedings of the 2017 ACM Workshop on Exploratory Search and Interactive Data Analytics. :69–74.

Labeled datasets are always limited, and oftentimes the quantity of labeled data is a bottleneck for data analytics. This especially affects supervised machine learning methods, which require labels for models to learn from the labeled data. Active learning algorithms have been proposed to help achieve good analytic models with limited labeling efforts, by determining which additional instance labels will be most beneficial for learning for a given model. Active learning is consistent with interactive analytics as it proceeds in a cycle in which the unlabeled data is automatically explored. However, in active learning users have no control of the instances to be labeled, and for text data, the annotation interface is usually document only. Both of these constraints seem to affect the performance of an active learning model. We hypothesize that visualization techniques, particularly interactive ones, will help to address these constraints. In this paper, we implement a pilot study of visualization in active learning for text classification, with an interactive labeling interface. We compare the results of three experiments. Early results indicate that visualization improves high-performance machine learning model building with an active learning algorithm.

Lai, J., Duan, B., Su, Y., Li, L., Yin, Q..  2017.  An active security defense strategy for wind farm based on automated decision. 2017 IEEE Power Energy Society General Meeting. :1–5.

With the development of smart grid, information and energy integrate deeply. For remote monitoring and cluster management, SCADA system of wind farm should be connected to Internet. However, communication security and operation risk put forward a challenge to data network of the wind farm. To address this problem, an active security defense strategy combined whitelist and security situation assessment is proposed. Firstly, the whitelist is designed by analyzing the legitimate packet of Modbus on communication of SCADA servers and PLCs. Then Knowledge Automation is applied to establish the Decision Requirements Diagram (DRD) for wind farm security. The D-S evidence theory is adopted to assess operation situation of wind farm and it together with whitelist offer the security decision for wind turbine. This strategy helps to eliminate the wind farm owners' security concerns of data networking, and improves the integrity of the cyber security defense for wind farm.

Phan, N., Wu, X., Hu, H., Dou, D..  2017.  Adaptive Laplace Mechanism: Differential Privacy Preservation in Deep Learning. 2017 IEEE International Conference on Data Mining (ICDM). :385–394.

In this paper, we focus on developing a novel mechanism to preserve differential privacy in deep neural networks, such that: (1) The privacy budget consumption is totally independent of the number of training steps; (2) It has the ability to adaptively inject noise into features based on the contribution of each to the output; and (3) It could be applied in a variety of different deep neural networks. To achieve this, we figure out a way to perturb affine transformations of neurons, and loss functions used in deep neural networks. In addition, our mechanism intentionally adds "more noise" into features which are "less relevant" to the model output, and vice-versa. Our theoretical analysis further derives the sensitivities and error bounds of our mechanism. Rigorous experiments conducted on MNIST and CIFAR-10 datasets show that our mechanism is highly effective and outperforms existing solutions.

Wang, Z., Wang, Y., Dong, B., Pracheta, S., Hamlen, K., Khan, L..  2020.  Adaptive Margin Based Deep Adversarial Metric Learning. 2020 IEEE 6th Intl Conference on Big Data Security on Cloud (BigDataSecurity), IEEE Intl Conference on High Performance and Smart Computing, (HPSC) and IEEE Intl Conference on Intelligent Data and Security (IDS). :100—108.

In the past decades, learning an effective distance metric between pairs of instances has played an important role in the classification and retrieval task, for example, the person identification or malware retrieval in the IoT service. The core motivation of recent efforts focus on improving the metric forms, and already showed promising results on the various applications. However, such models often fail to produce a reliable metric on the ambiguous test set. It happens mainly due to the sampling process of the training set, which is not representative of the distribution of the negative samples, especially the examples that are closer to the boundary of different categories (also called hard negative samples). In this paper, we focus on addressing such problems and propose an adaptive margin deep adversarial metric learning (AMDAML) framework. It exploits numerous common negative samples to generate potential hard (adversarial) negatives and applies them to facilitate robust metric learning. Apart from the previous approaches that typically depend on the search or data augmentation to find hard negative samples, the generation of adversarial negative instances could avoid the limitation of domain knowledge and constraint pairs' amount. Specifically, in order to prevent over fitting or underfitting during the training step, we propose an adaptive margin loss that preserves a flexible margin between the negative (include the adversarial and original) and positive samples. We simultaneously train both the adversarial negative generator and conventional metric objective in an adversarial manner and learn the feature representations that are more precise and robust. The experimental results on practical data sets clearly demonstrate the superiority of AMDAML to representative state-of-the-art metric learning models.