As data mining matures as a field and develops more powerful algorithms for discovering and exploiting patterns in data, the amount of data about individuals that is collected and stored continues to rapidly increase. This increase in data heightens concerns that data mining violates individual privacy. The goal of data mining is to derive aggregate conclusions, which should not reveal sensitive information. However, the data-mining algorithms run on databases containing information about individuals which may be sensitive. The goal of privacy-preserving data mining is to provide high-quality aggregate conclusions while protecting the privacy of the constituent individuals. The field of "privacy-preserving data mining" encompasses a wide variety of different techniques and approaches, and considers many different threat and trust models. Some techniques use perturbation, where noise is added (either directly to the database that is the input to the algorithm or to the output of queries) to obscure values of sensitive attributes; some use generalization, where identifying attributes are given less specific values; and some use cryp- tography, where joint computations between multiple parties are performed on encrypted data to hide inputs. Because these approaches are applied to different scenarios with different threat models, their overall e ectiveness and privacy properties are incomparable. In this thesis I take a pragmatic approach to privacy-preserving data mining and attempt to determine which techniques are suitable to real-world problems that a data miner might wish to solve, such as evaluating and learning decision-tree classifiers. I show that popular techniques for sanitizing databases prior to publication either fail to provide any meaningful privacy guarantees, or else degrade the data to the point of having only negligible data-mining utility. Cryptographic techniques for secure multi-party computation are a natural alternative to sanitized data publication, and guarantee the privacy of inputs by performing computations on encrypted data. Because of its heavy reliance on public-key cryptography, it is conventionally thought to be too slow to apply to real-world problems. I show that tailor-made protocols for specific data-mining problems can be made fast enough to run on real-world problems, and I strengthen this claim with empirical runtime analysis using prototype implementations. I also expand the use of secure computation beyond its traditional scope of applying a known algorithm to private inputs by showing how it can be used to e ciently apply a private algorithm, chosen from a specific class of algorithms, to a private input. / text

There are two main approaches for intrusion detection: signature-based and anomaly-based. Signature-based detection employs pattern matching to match attack signatures with observed data making it ideal for detecting known attacks. However, it cannot detect unknown attacks for which there is no signature available. Anomaly-based detection builds a profile of normal system behavior to detect known and unknown attacks as behavioral deviations. However, it has a drawback of a high false alarm rate. In this thesis, we describe our anomaly-based IDS designed for detecting intrusions in cryptographic and application-level protocols. Our system has several unique characteristics, such as the ability to monitor cryptographic protocols and application-level protocols embedded in encrypted sessions, a very lightweight monitoring process, and the ability to react to protocol misuse by modifying protocol response directly.

In the last few years, security has become one of the key challenges in computing systems. Failures in the secure operations of these systems have led to massive information leaks and cyber-attacks. Case in point, the identity leaks from Equifax in 2016, Spectre and Meltdown attacks to Intel and AMD processors in 2017, Cyber-attacks on Facebook in 2018. These recent attacks have shown that the intruders attack different layers of the systems, from low-level hardware to software as a service(SaaS). To protect the systems, the defense mechanisms should confront the attacks in the different layers of the systems. In this work, we propose four security mechanisms for computing systems: (i ) using backside imaging to detect Hardware Trojans (HTs) in Application Specific Integrated Circuits (ASICs) chips, (ii ) developing energy-efficient reconfigurable cryptographic engines, (iii) examining the feasibility of malware detection using Hardware Performance Counters (HPC). Most of the threat models assume that the root of trust is the hardware running beneath the software stack. However, attackers can insert malicious hardware blocks, i.e. HTs, into the Integrated Circuits (ICs) that provide back-doors to the attackers or leak confidential information. HTs inserted during fabrication are extremely hard to detect since their overheads in performance and power are below the variations in the performance and power caused by manufacturing. In our work, we have developed an optical method that identifies modified or replaced gates in the ICs. We use the near-infrared light to image the ICs because silicon is transparent to near-infrared light and metal reflects infrared light. We leverage the near-infrared imaging to identify the locations of each gate, based on the signatures of metal structures reflected by the lowest metal layer. By comparing the imaged results to the pre-fabrication design, we can identify any modifications, shifts or replacements in the circuits to detect HTs. With the trust of the silicon, the computing system must use secure communication channels for its applications. The low-energy cost devices, such as the Internet of Things (IoT), leverage strong cryptographic algorithms (e.g. AES, RSA, and SHA) during communications. The cryptographic operations cause the IoT devices a significant amount of power. As a result, the power budget limits their applications. To mitigate the high power consumption, modern processors embed these cryptographic operations into hardware primitives. This also improves system performance. The hardware unit embedded into the processor provides high energy-efficiency, low energy cost. However, hardware implementations limit flexibility. The longevity of theIoTs can exceed the lifetime of the cryptographic algorithms. The replacement of the IoT devices is costly and sometimes prohibitive, e.g., monitors in nuclear reactors.In order to reconfigure cryptographic algorithms into hardware, we have developed a system with a reconfigurable encryption engine on the Zedboard platform. The hardware implementation of the engine ensures fast, energy-efficient cryptographic operations. With reliable hardware and secure communication channels in place, the computing systems should detect any malicious behaviors in the processes. We have explored the use of the Hardware Performance Counters (HPCs) in malware detection. HPCs are hardware units that count micro-architectural events, such as cache hits/misses and floating point operations. Anti-virus software is commonly used to detect malware but it also introduces performance overhead. To reduce anti-virus performance overhead, many researchers propose to use HPCs with machine learning models in malware detection. However, it is counter-intuitive that the high-level program behaviors can manifest themselves in low-level statics. We perform experiments using 2 ∼ 3 × larger program counts than the previous works and perform a rigorous analysis to determine whether HPCs can be used to detect malware. Our results show that the False Discovery Rate of malware detection can reach 20%. If we deploy this detection system on a fresh installed Windows 7 systems, among 1,323 binaries, 198 binaries would be flagged as malware.

<p>With the escalating growth of e-commerce in today’s society, many e-commerce sites have emerged that offer products on the Internet. To be able to verify orders from customers, some sites require sensitive information from their customers such as credit card details that is stored in their databases. The security of these sites has become the concern of many and it is a common opinion among the public that such sites cannot be trusted.</p><p>The FiLDB architecture presents an interesting approach for increasing the security of Internet connected databases. This approach is, in short, based on firewall protection; one external firewall protecting an external network, which in turn is connected to an internal network, which was protected by a internal firewall. A database is kept on each network. There are however few issues that are unsolved in the FiLDB architecture. One of them and the problem that is covered in this report is how a user could securely insert, modify and fetch sensitive data into the internal database which stores the sensitive data.</p><p>In this work a few selected cryptographic protocols are studied by evaluating them with respect to five security criteria: confidentiality, authentication, integrity, key management and nonrepudiation. The initial selection of cryptographic protocols is mainly based on applicability in e-commerce systems. Based on the evaluation, one of the protocols is chosen to be implemented with the FiLDB architecture and then the extended architecture was evaluated.</p><p>This project shows that, by integrating a security protocol into architectures such as the FiLDB, the security of the system can be increased substantially.</p>

Σκοπός αυτής της διπλωαμτικής εργασίας είναι ο σχεδιασμός και η υλοποίηση (σε γλώσσα περιγραφής υλικού VHDL) των κρυπτογραφικών αλγορίθμων SHA-1 και SHA-256. Απώτερος στόχος υπήρξε η σύγκριση και αξιολόγηση της κάθε υλοποίησης για να επιτευχθεί όσο το δυνατόν καλύτερο throughput/area. Οι τεχνικές που χρησιμοποιήθηκαν ήταν η βελτίωση του αλγορίθμου μαζί με την τεχνική της διοχέτευσης. / Aim of this work is the planning and the concretisation (in VHDL)of the cryptographic algorithms SHA-1 and SHA-256. Final objective is the comparison and evaluation of each concretisation in order to achieve optimum throughput/area. The techniques that were used were the improvement of algorithm with the technique of pipeline.

With the escalating growth of e-commerce in today’s society, many e-commerce sites have emerged that offer products on the Internet. To be able to verify orders from customers, some sites require sensitive information from their customers such as credit card details that is stored in their databases. The security of these sites has become the concern of many and it is a common opinion among the public that such sites cannot be trusted. The FiLDB architecture presents an interesting approach for increasing the security of Internet connected databases. This approach is, in short, based on firewall protection; one external firewall protecting an external network, which in turn is connected to an internal network, which was protected by a internal firewall. A database is kept on each network. There are however few issues that are unsolved in the FiLDB architecture. One of them and the problem that is covered in this report is how a user could securely insert, modify and fetch sensitive data into the internal database which stores the sensitive data. In this work a few selected cryptographic protocols are studied by evaluating them with respect to five security criteria: confidentiality, authentication, integrity, key management and nonrepudiation. The initial selection of cryptographic protocols is mainly based on applicability in e-commerce systems. Based on the evaluation, one of the protocols is chosen to be implemented with the FiLDB architecture and then the extended architecture was evaluated. This project shows that, by integrating a security protocol into architectures such as the FiLDB, the security of the system can be increased substantially.

Cette these concerne la verification formelle et la composition de protocoles de securite, motivees en particulier par l'analyse des protocoles de vote electronique. Les chapitres 3 a 5 ont comme sujet la verification de protocoles de securite et le Chapitre 6 vise la composition.Nous montrons dans le Chapitre 3 comment reduire certains problemes d'une algebre quotient des termes a l'algebre libre des termes en utilisant des ensembles fortement complets de variants. Nous montrons que, si l'algebre quotient est donnee par un systeme de reecriture de termes convergent et optimalement reducteur (optimally reducing), alors des ensembles fortement complets de variants existent et sont finis et calculables.Dans le Chapitre 4, nous montrons que l'equivalence statique pour (des classes) de theories equationnelles, dont les theories sous-terme convergentes, la theorie de l'engagement a trappe (trapdoor commitment) et la theorie de signature en aveugle (blind signatures), est decidable en temps polynomial. Nous avons implemente de maniere efficace cette procedure.Dans le Chapitre 5, nous etendons la procedure de decision precedente a l'equivalence de traces. Nous utilisons des ensembles fortement complets de variants du Chapitre 3 pour reduire le probleme a l'algebre libre. Nous modelisons chaque trace du protocole comme une theorie de Horn et nous utilisons un raffinement de la resolution pour resoudre cette theorie. Meme si nous n'avons pas reussi a prouver que la procedure de resolution termine toujours, nous l'avons implementee et utilisee pour donner la premiere preuve automatique de l'anonymat dans le protocole de vote electronique FOO.Dans le Chapitre 6, nous etudions la composition de protocoles. Nous montrons que la composition de deux protocoles qui utilisent des primitives cryptographiques disjointes est sure s'ils ne revelent et ne reutilisent pas les secrets partages. Nous montrons qu'une forme d'etiquettage de protocoles est suffisante pour assurer la disjonction pour un ensemble fixe de primitives cryptographiques. / This thesis is about the formal verification and composition of security protocols, motivated by applications to electronic voting protocols. Chapters 3 to 5 concern the verification of security protocols while Chapter 6 concerns composition.We show in Chapter 3 how to reduce certain problems from a quotient term algebra to the free term algebra via the use of strongly complete sets of variants. We show that, when the quotient algebra is given by a convergent optimally reducing rewrite system, finite strongly complete sets of variants exist and are effectively computable.In Chapter 4, we show that static equivalence for (classes of) equational theories including subterm convergent equational theories, trapdoor commitment and blind signatures is decidable in polynomial time. We also provide an efficient implementation.In Chapter 5 we extend the previous decision procedure to handle trace equivalence. We use finite strongly complete sets of variants introduced in Chapter 3 to get rid of the equational theory and we model each protocol trace as a Horn theory which we solve using a refinement of resolution. Although we have not been able to prove that this procedure always terminates, we have implemented it and used it to provide the first automated proof of vote privacy of the FOO electronic voting protocol.In Chapter 6, we study composition of protocols. We show that two protocols that use arbitrary disjoint cryptographic primitives compose securely if they do not reveal or reuse any shared secret. We also show that a form of tagging is sufficient to provide disjointness in the case of a fixed set of cryptographic primitives.

The purpose of this work was first to describe the requirements for cryptographic protocols. Furthermore, the classification of these protocols should have been made with specific examples given. The aim of the next part of the work was to describe the methods which are suitable for description and modeling of cryptographic protocols. This work also addressed the analysis of cryptographic protocols by appropriate analytical means. The CSP method for modeling of the cryptographic protocols was applied in the practical part. The Yahalom protocol was selected as a protocol suitable for modeling. Two analysis was made. The first analysis concerned the standard version of the Yahalom protocol, which was tested to the requirements of cryptographic properties of the secrecy and authenticity. The second analysis was based on the possibility of disclosure of the key, including counterexamples and traces given by FDR. The first analysis did not reveal any weakening, in terms of two cryptographic properties. To demonstrate the possibility of FDR, Yahalom protocol was modified in order to cause the situation when the disclosure of keys appears. FDR then finds the exact procedure that an intruder must make to get the possession of the key.

Cryptographic accumulators are a tool for compact set representation and secure set membership proofs. When an element is added to a set by means of an accumulator, a membership witness is generated. This witness can later be used to prove the membership of the element. Typically, the membership witness has to be synchronized with the accumulator value, and to be updated every time another element is added to the accumulator. In this work we propose an accumulator that, unlike any prior scheme, does not require strict synchronization. In our construction a membership witness needs to be updated only a logarithmic number of times in the number of subsequent element additions. Thus, an out-of-date witness can be easily made current. Vice versa, a verifier with an out-of-date accumulator value can still verify a current membership witness. These properties make our accumulator construction uniquely suited for use in distributed applications, such as blockchain-based public key infrastructures.

Many studies show that manual code generation is error-prone and results in vulnerabilities. Vulnerability fixing has been shown as the most time-consuming process among multiple steps of code repair. To help developers repair these security vulnerabilities, my dissertation aims to develop an automatic or semi-automatic secure code generation system with neural network based approaches. Trained with huge amounts of good-quality code, I expect the neural network to learn the secure usage and produce the correct code suggestions. Despite the great success of neural networks, the vision of comprehending and generating programming languages through neural networks has not been fully realized. There are many fundamental questions that need to be answered. These questions include 1) what are the accuracy impacts of the various choices in code embedding? 2) How to address the accuracy challenges caused by the programming language specific properties in the task of secure code suggestion? My dissertation work answers the two questions with a systematical measurement study and specialized neural network designs. My experiments show that program analysis is a necessary preprocessing step to guide the code embedding – resulting in a 36.1% accuracy improvement. Furthermore, I identify two previously unreported deficiencies in the cryptographic API suggestion task. To close the gap, I invent a highly accurate API method suggestion solution, referred to as Multi-HyLSTM, with specialized neural network designs to recognize unique programming language characteristics. My work points out the important differences between natural languages and programming languages, which pure data-driven learning approaches may not recognize. / Doctor of Philosophy / Neural network techniques that automatically learn rules from data show great potential to provide vulnerability-agnostic solutions for securing code. Recent research community has witnessed the rapid progress of neural network techniques in various application domains, such as computer vision, natural language processing, etc. However, how to harness the success of neural network based approaches for dealing with programs is still largely unknown. Many fundamental questions are required to be answered. This dissertation aims to provide neural network based solutions to help developers write secure code, as well as answer several important but unknown research questions about promoting neural network based approaches specialized for the programming language domain. Learning from Java cryptographic code, I explore the accuracy challenges for neural networks to understand the secure API usage rules and generate appropriate suggestions based on them. One of my research focuses is on how to express code in a way that neural networks can comprehend, aka code embedding. Code embedding is the process of transforming code into numeric vectors. It is important for accuracy as all the subsequent neural network calculation is performed on it. I conduct a systematic comparison to evaluate several key embedding design choices and reveal their impacts on accuracy improvements. To further improve the accuracy, I focus on the accuracy challenges in the specific task, generating API suggestions by neural networks. I identify the unreported program dependency specific challenges and present several specialized neural network designs to address them.