Evasion attacks pose a substantial risk to the application of Machine Learning (ML) in Cybersecurity, potentially leading to safety hazards or security breaches in large-scale deployments. Adversaries can employ evasion attacks as an initial tactic to deceive malware or network scanners using ML, thereby orchestrating traditional cyber attacks to disrupt systems availability or compromise integrity. Adversarial data designed to fool AI systems for cybersecurity can be engineered by strategically selecting, modifying, or creating test instances. This paper presents novel defender-centric techniques and metrics for mitigating evasion attacks by leveraging adversarial knowledge, exploring potential exploitation methods, and enhancing alarm detection capabilities. We first introduce two new evasion resistance metrics: adversarial failure rate ( afr ) and adversarial failure curves ( afc ). These metrics generalize previous approaches, as they can be applied to threshold classifiers, facilitating analyses for adversarial attacks comparable to those performed with Receiver Operating Characteristics (ROC) curve. Subsequently, we propose two novel evasion resistance techniques (trainset size pinning and model matrix), extending research in keyed intrusion detection and randomization. We explore the application of proposed techniques and metrics to an intrusion detection system as a pilot study using two public datasets, ‘BETH 2021’ and ‘Kyoto 2015’, which are well-established cybersecurity datasets for uncertainty and robustness benchmarking. The experimental results demonstrate that the combination of the proposed randomization techniques consistently produces remarkable improvement over other known randomization techniques.

Several Mobile Device Management (MDM) and Mobile Application Management (MAM) services have been launched on the market. However, these services suffer from two important limitations: reduced granularity and need for app developers to include third party SDKs. We present AppBox, a novel black-box app-sandboxing solution for app customisation for stock Android devices. AppBox enables enterprises to select any app, even highly-obfuscated, from any market and perform a set of target customisations by means of fine-grained security policies. We have implemented and tested AppBox on various smartphones and Android versions. The evaluation shows that AppBox can effectively enforce fine-grained policies on a wide set of existing apps, with an acceptable overhead.

Identity Management systems come with a promise of simpler, centralized, and more secure handling of user data, credentials and authorizations. Service providers can thus be separated from an identity provider (IdP), and users will benefit from single sign on mechanisms. However, identity providers become single points of failure, from a security and trust perspective. In particular, in this paper, we address the protection of user privacy against IdPs profiling. The IdP should not be aware of which services the user will access, or about the details of such service use. This paper discusses the difficulties of this type of privacy protection problem and surveys existing solutions. We then identify a novel solution and improve an existing one: (1) moving part of the access management logic to the client, via a Web Browser extension, and (2) elaborating on previously proposed solutions based on the concept of a linking service, i.e. a server separated from the IdP and from the Service Provider, that will perform some of the authentication steps. Proof of the principle implementations of both solutions are made available to the user.

Unlike other access control systems based on biometric features, keystroke analysis has not led to techniques providing an acceptable level of accuracy. The reason is probably the intrinsic variability of typing dynamics, versus other---very stable---biometric characteristics, such as face or fingerprint patterns. In this paper we present an original measure for keystroke dynamics that limits the instability of this biometric feature. We have tested our approach on 154 individuals, achieving a False Alarm Rate of about 4% and an Impostor Pass Rate of less than 0.01%. This performance is reached using the same sampling text for all the individuals, allowing typing errors, without any specific tailoring of the authentication system with respect to the available set of typing samples and users, and collecting the samples over a 28.8-Kbaud remote modem connection.

We study the learnability of multiplicity automata in Angluin's exact learning model , and we investigate its applications. Our starting point is a known theorem from automata theory relating the number of states in a minimal multiplicity automaton for a function to the rank of its Hankel matrix. With this theorem in hand, we present a new simple algorithm for learning multiplicity automata with improved time and query complexity, and we prove the learnability of various concept classes. These include (among others): -The class of disjoint DNF, and more generally satisfy- O (1) DNF. -The class of polynomials over finite fields. -The class of bounded-degree polynomials over infinite fields. -The class of XOR of terms. -Certain classes of boxes in high dimensions. In addition, we obtain the best query complexity for several classes known to be learnable by other methods such as decision trees and polynomials over GF(2). While multiplicity automata are shown to be useful to prove the learnability of some subclasses of DNF formulae and various other classes, we study the limitations of this method. We prove that this method cannot be used to resolve the learnability of some other open problems such as the learnability of general DNF formulas or even k -term DNF for k = ω(log n ) or satisfy- s DNF formulas for s = ω(1). These results are proven by exhibiting functions in the above classes that require multiplicity automata with super-polynomial number of states.

The important problem of user password selection is addressed and a new proactive password-checking technique is presented. In a training phase, a decision tree is generated based on a given dictionary of weak passwords. Then, the decision tree is used to determine whether a user password should be accepted. Experimental results described here show that the method leads to a very high dictionary compression (up to 1000 to 1) with low error rates (of the order of 1%). A prototype implementation, called ProCheck, is made available online. We survey previous approaches to proactive password checking, and provide an in-depth comparison.

The important problem of user password selection is addressed and a new proactive password checking technique is presented.In a training phase, a decision tree is generated based on a given dictionary of weak passwords.Then, the decision tree is used to determine whether a user password should be accepted.Experimental results described here show that the method leads to very high dictionary compression (from 100 to 3 in the average) with low error rates (of the order of 1%).We survey previous approaches to proactive password checking, and provide an in-depth comparison.

A framework for secure WWW client/server communication is proposed. Strong end-to-end encryption and authentication is achieved by means of public key techniques. A particular certification infrastructure is developed that helps assign responsibilities in case of disputes. Such issues are increasin gly important in WWW applications and are not dealt with in a satisfactory way by current certification schemes. Actual communication is done with the HTTP protocol unchanged and by using standard commercial browsers, because widespread usability is a goal. Encryption and authentication is done separately based on the execution of applets running on the client machine.

The learnability of multiplicity automata has attracted a lot of attention, mainly because of its implications on the learnability of several classes of DNF formulae. The authors further study the learnability of multiplicity automata. The starting point is a known theorem from automata theory relating the number of states in a minimal multiplicity automaton for a function f to the rank of a certain matrix F. With this theorem in hand they obtain the following results: a new simple algorithm for learning multiplicity automata with a better query complexity. As a result, they improve the complexity for all classes that use the algorithms of Bergadano and Varricchio (1994) and Ohnishi et al. (1994) and also obtain the best query complexity for several classes known to be learnable by other methods such as decision trees and polynomials over GF(2). They prove the learnability of some new classes that were not known to be learnable before. Most notably, the class of polynomials over finite fields, the class of bounded-degree polynomials over infinite fields, the class of XOR of terms, and a certain class of decision trees. While multiplicity automata were shown to be useful to prove the learnability of some subclasses of DNF formulae and various other classes, they study the limitations of this method. They prove that this method cannot be used to resolve the learnability of some other open problems such as the learnability of general DNF formulae or even K-term DNF for k=/spl omega/ (log n) or satisfy-s DNF formulae for s=/spl omega/(1). These results are proven by exhibiting functions in the above classes that require multiplicity automata with superpolynomial number of states.

We consider the problem of identifying the behavior of an unknown automaton with multiplicity in the field Q of rational numbers (Q-automaton) from multiplicity and equivalence queries. We provide an algorithm which is polynomial in the size of the Q-automaton and in the maximum length of the given counterexamples. As a consequence, we have that Q-automata are probably approximately correctly learnable (PAC-learnable) in polynomial time when multiplicity queries are allowed. A corollary of this result is that regular languages are polynomially predictable using membership queries with respect to the representation of unambiguous nondeterministic automata. This is important since there are unambiguous automata such that the equivalent deterministic automaton has an exponentially larger number of states.

Given a program P and a set of alternative programs //', we generate a sequence of test cases that are adequate, in the sense that they distinguish the given program from all alternatives The, m(,thod is related to fault-based approaches to test case generation, but programs in P need not he s]mp]e mutations of P. The technique for generating an adequate test set is based on the inductive learning of programs from finite sets of input-output examples: given a partial test set.we generate inductively a program P' E P which is consistent with P on those input values; then we look for an input value that distinguishes P from P', and we repeat the process until no program except P can be induced from the generated examples.We show that the obtained test set is adequate with respect to the alternatives belonging to P. The method IS made possible by a program induction procedure which has evolved from recent research in mnchine Iei]rnlng and inductive logic programming.An implemented version of the test case ~encration procedure is demonstrated on simple and more complex list-processing programs.and tb(, scalability of' the approach is discussed, ('ate~ories and Subject Descriptors: D.2.

An efficient method for learning (trapezoidal) membership functions for fuzzy predicates is presented. Positive and negative examples of one class are given together with a system of classification rules. The learned membership functions can be used for the fuzzy predicates occurring in the given rules to classify further examples. We show that the obtained classification is approximately correct with high probability. This justifies the obtained fuzzy sets within one particular classification problem, instead of relying on a subjective meaning of fuzzy predicates as normally done by a domain expert.

Given a program P and a set of alternative programs P, we generate a sequence of test cases that are adequate, in the sense that they distinguish the given program from all alternatives. The method is related to fault-based approaches to program testing, but programs in P need not be simple mutations of P. The technique for generating an adequate test set is based on the inductive learning of programs from finite sets of input-output examples: given a partial test set, we generate inductively a program P'E P which is consistent with P on those input values; then we look for an input value that distinguishes P from P', and repeat the process until no program except P can be induced from the generated examples. We show that the so obtained test set is adequate w.r.t. the alternatives belonging to P. The method is made possible by a practical program induction procedure, which has evolved from recent research in Machine Learning and Inductive Logic Programming.

As real logic programmers normally use cut (!), an effective learning procedure for logic programs should be able to deal with it. Because the cut predicate has only a procedural meaning, clauses containing cut cannot be learned using an extensional evaluation method, as is done in most learning systems. On the other hand, searching a space of possible programs (instead of a space of independent clauses) is unfeasible. An alternative solution is to generate first a candidate base program which covers the positive examples, and then make it consistent by inserting cut where appropriate. The problem of learning programs with cut has not been investigated before and this seems to be a natural and reasonable approach. We generalize this scheme and investigate the difficulties that arise. Some of the major shortcomings are actually caused, in general, by the need for intensional evaluation. As a conclusion, the analysis of this paper suggests, on precise and technical grounds, that learning cut is difficult, and current induction techniques should probably be restricted to purely declarative logic languages.

The computational learning approach shows that the concept descriptions acquired from examples are approximately correct with a degree of probability that grows with the size of the training sample. The same problem has also been widely investigated in the field of pattern recognition under a variety of problem settings. Some of the results obtained in both fields are surveyed and compared, and the limits of their applicability are analyzed. Moreover, new and tighter bounds for the growth function of some classes of Boolean formulas are presented.>

The concept of an inductive relation is introduced, as a natural development of other forms of intentional information, such as views and relations defined deductively. A class of top-down methods for computing such inductive relations is analyzed. Major problems produced by recursive and interdependent relations are considered.>

The results of extensive experimentation aimed at assessing the concrete possibilities of automatically building a diagnostic expert system, to be used in-field in an industrial domain, by means of machine learning techniques, are described. The system, ENIGMA, is an incremental version of the ML-SMART system, which acquires a network of first-order logic rules, starting from a set of classified examples and a domain theory. An application is described that consists of discovering malfunctions in electromechanical apparatus. ENIGMA's efficacy in acquiring sophisticated knowledge and handling complex structured examples is largely due to its underlying database management system, which supports the learning operators, defined at the abstract level, with a set of primitives, taken from the field of deductive databases. An expert system, MEPS, devoted to the same task, has also been manually developed. A number of comparisons along different dimensions of the manual and automatic development process have been possible, allowing some practical indications to be suggested.>

This article presents a knowledge based methodology for recognizing concept instances in complex data such as natural scenes or speech signals. the architecture of a prototype system performing this task is also described. In order to obtain the capability of handling noisy patterns, the knowledge representation is based on continuous valued logics, and the inference engine is able to do sophisticated reasoning about the knowledge it uses in order to take into account the possible degradation of cues in the pattern. The knowledge base is subdivided into a bulk knowledge and a degradation theory. the bulk knowledge consists of production rules capturing discriminating cues in the signal as they appear in the normal cases. the degradation theory describes how the cues can be degraded owing to insertion and deletion errors. The whole classification process consists of two phases. First of all, the rules of the bulk knowledge are applied to the pattern, trying to obtain a suitable classification, and the evidence assigned to the rules is combined with a technique based on Dempster-Shafer theory. In the second phase, the classification is refined taking into account also the possible degradations in the pattern. the resulting system is then able to deal with different kinds of uncertainty and errors, such as fluctuation in the values measured for the pattern and insertion-deletion errors. the method is illustrated and evaluated on a simple example of spoken word recognition.

A system that performs automated concept acquisition from examples and has been specially designed to work in noisy environments is presented. The learning methodology is aimed at the target problem of finding discriminant descriptions of a given set of concepts and uses both examples and counterexamples. The learned knowledge is expressed in the form of production rules, organized into separate clusters, linked together in a graph structure. Knowledge extraction is guided by a top-down control strategy, through a process of specialization. The system also utilizes a technique of problem reduction to contain the computational complexity. Several criteria are proposed for evaluating the acquired knowledge. The methodology has been tested on a problem in the field of speech recognition and the experimental results obtained are reported and discussed.>

This paper surveys a long term project, aimed at providing a general methodology for building up and maintaining an expert system oriented to Pattern Recognition problems. The methodology makes use of an integrated set of modules, performing different functions but sharing a common knowledge representation scheme. In particular, a learning module allows to acquire the knowledge automatically from a set of examples and another module performs sophisticated reasoning, on the basis of the available knowledge, during the recognition phase.

In this paper a prototype Expert System oriented to signal and pattern analysis is described together with a general methodology based on a hypothesize-and-test strategy similar to the one used by a human expert. This paper focuses on the knowledge base architecture and on its use. In order to make it capable of dealing with noisy patterns, the knowledge description is based on Fuzzy logic and the inference engine is able to reason about the knowledge it uses. The system is being applied to the phonetic analysis of the voice signal in speech recognition, as a case study.