Do Large Language Models Show Biases in Causal Learning?

Causal learning is the cognitive process of developing the capability of making causal inferences based on available information, often guided by normative principles. This process is prone to errors and biases, such as the illusion of causality, in which people perceive a causal relationship between two variables despite lacking supporting evidence. This cognitive bias has been proposed to underlie many societal problems, including social prejudice, stereotype formation, misinformation, and superstitious thinking. In this research, we investigate whether large language models (LLMs) develop causal illusions, both in real-world and controlled laboratory contexts of causal learning and inference. To this end, we built a dataset of over 2K samples including purely correlational cases, situations with null contingency, and cases where temporal information excludes the possibility of causality by placing the potential effect before the cause. We then prompted the models to make statements or answer causal questions to evaluate their tendencies to infer causation erroneously in these structured settings. Our findings show a strong presence of causal illusion bias in LLMs. Specifically, in open-ended generation tasks involving spurious correlations, the models displayed bias at levels comparable to, or even lower than, those observed in similar studies on human subjects. However, when faced with null-contingency scenarios or temporal cues that negate causal relationships, where it was required to respond on a 0-100 scale, the models exhibited significantly higher bias. These findings suggest that the models have not uniformly, consistently, or reliably internalized the normative principles essential for accurate causal learning.

Geospatial Trajectory Generation via Efficient Abduction: Deployment for Independent Testing

The ability to generate artificial human movement patterns while meeting location and time constraints is an important problem in the security community, particularly as it enables the study of the analog problem of detecting such patterns while maintaining privacy. We frame this problem as an instance of abduction guided by a novel parsimony function represented as an aggregate truth value over an annotated logic program. This approach has the added benefit of affording explainability to an analyst user. By showing that any subset of such a program can provide a lower bound on this parsimony requirement, we are able to abduce movement trajectories efficiently through an informed (i.e., A*) search. We describe how our implementation was enhanced with the application of multiple techniques in order to be scaled and integrated with a cloud-based software stack that included bottom-up rule learning, geolocated knowledge graph retrieval/management, and interfaces with government systems for independently conducted government-run tests for which we provide results. We also report on our own experiments showing that we not only provide exact results but also scale to very large scenarios and provide realistic agent trajectories that can go undetected by machine learning anomaly detectors.

Extensions to Generalized Annotated Logic and an Equivalent Neural Architecture

While deep neural networks have led to major advances in image recognition, language translation, data mining, and game playing, there are well-known limits to the paradigm such as lack of explainability, difficulty of incorporating prior knowledge, and modularity. Neuro symbolic hybrid systems have recently emerged as a straightforward way to extend deep neural networks by incorporating ideas from symbolic reasoning such as computational logic. In this paper, we propose a list desirable criteria for neuro symbolic systems and examine how some of the existing approaches address these criteria. We then propose an extension to generalized annotated logic that allows for the creation of an equivalent neural architecture comprising an alternate neuro symbolic hybrid. However, unlike previous approaches that rely on continuous optimization for the training process, our framework is designed as a binarized neural network that uses discrete optimization. We provide proofs of correctness and discuss several of the challenges that must be overcome to realize this framework in an implemented system.

Reasoning about Complex Networks: A Logic Programming Approach

Reasoning about complex networks has in recent years become an important topic of study due to its many applications: the adoption of commercial products, spread of disease, the diffusion of an idea, etc. In this paper, we present the MANCaLog language, a formalism based on logic programming that satisfies a set of desiderata proposed in previous work as recommendations for the development of approaches to reasoning in complex networks. To the best of our knowledge, this is the first formalism that satisfies all such criteria. We first focus on algorithms for finding minimal models (on which multi-attribute analysis can be done), and then on how this formalism can be applied in certain real world scenarios. Towards this end, we study the problem of deciding group membership in social networks: given a social network and a set of groups where group membership of only some of the individuals in the network is known, we wish to determine a degree of membership for the remaining group-individual pairs. We develop a prototype implementation that we use to obtain experimental results on two real world datasets, including a current social network of criminal gangs in a major U.S.\ city. We then show how the assignment of degree of membership to nodes in this case allows for a better understanding of the criminal gang problem when combined with other social network mining techniques -- including detection of sub-groups and identification of core group members -- which would not be possible without further identification of additional group members.

On the Importance of Domain-specific Explanations in AI-based Cybersecurity Systems (Technical Report)

With the availability of large datasets and ever-increasing computing power, there has been a growing use of data-driven artificial intelligence systems, which have shown their potential for successful application in diverse areas. However, many of these systems are not able to provide information about the rationale behind their decisions to their users. Lack of understanding of such decisions can be a major drawback, especially in critical domains such as those related to cybersecurity. In light of this problem, in this paper we make three contributions: (i) proposal and discussion of desiderata for the explanation of outputs generated by AI-based cybersecurity systems; (ii) a comparative analysis of approaches in the literature on Explainable Artificial Intelligence (XAI) under the lens of both our desiderata and further dimensions that are typically used for examining XAI approaches; and (iii) a general architecture that can serve as a roadmap for guiding research efforts towards the development of explainable AI-based cybersecurity systems -- at its core, this roadmap proposes combinations of several research lines in a novel way towards tackling the unique challenges that arise in this context.

An Approach to Characterize Graded Entailment of Arguments through a Label-based Framework

Argumentation theory is a powerful paradigm that formalizes a type of commonsense reasoning that aims to simulate the human ability to resolve a specific problem in an intelligent manner. A classical argumentation process takes into account only the properties related to the intrinsic logical soundness of an argument in order to determine its acceptability status. However, these properties are not always the only ones that matter to establish the argument's acceptability---there exist other qualities, such as strength, weight, social votes, trust degree, relevance level, and certainty degree, among others.

DARKMENTION: A Deployed System to Predict Enterprise-Targeted External Cyberattacks

Recent incidents of data breaches call for organizations to proactively identify cyber attacks on their systems. Darkweb/Deepweb (D2web) forums and marketplaces provide environments where hackers anonymously discuss existing vulnerabilities and commercialize malicious software to exploit those vulnerabilities. These platforms offer security practitioners a threat intelligence environment that allows to mine for patterns related to organization-targeted cyber attacks. In this paper, we describe a system (called DARKMENTION) that learns association rules correlating indicators of attacks from D2web to real-world cyber incidents. Using the learned rules, DARKMENTION generates and submits warnings to a Security Operations Center (SOC) prior to attacks. Our goal was to design a system that automatically generates enterprise-targeted warnings that are timely, actionable, accurate, and transparent. We show that DARKMENTION meets our goal. In particular, we show that it outperforms baseline systems that attempt to generate warnings of cyber attacks related to two enterprises with an average increase in F1 score of about 45% and 57%. Additionally, DARKMENTION was deployed as part of a larger system that is built under a contract with the IARPA Cyber-attack Automated Unconventional Sensor Environment (CAUSE) program. It is actively producing warnings that precede attacks by an average of 3 days.

Argumentation Models for Cyber Attribution

A major challenge in cyber-threat analysis is combining information from different sources to find the person or the group responsible for the cyber-attack. It is one of the most important technical and policy challenges in cyber-security. The lack of ground truth for an individual responsible for an attack has limited previous studies. In this paper, we take a first step towards overcoming this limitation by building a dataset from the capture-the-flag event held at DEFCON, and propose an argumentation model based on a formal reasoning framework called DeLP (Defeasible Logic Programming) designed to aid an analyst in attributing a cyber-attack. We build models from latent variables to reduce the search space of culprits (attackers), and show that this reduction significantly improves the performance of classification-based approaches from 37% to 62% in identifying the attacker.

An Argumentation-Based Framework to Address the Attribution Problem in Cyber-Warfare

Attributing a cyber-operation through the use of multiple pieces of technical evidence (i.e., malware reverse-engineering and source tracking) and conventional intelligence sources (i.e., human or signals intelligence) is a difficult problem not only due to the effort required to obtain evidence, but the ease with which an adversary can plant false evidence. In this paper, we introduce a formal reasoning system called the InCA (Intelligent Cyber Attribution) framework that is designed to aid an analyst in the attribution of a cyber-operation even when the available information is conflicting and/or uncertain. Our approach combines argumentation-based reasoning, logic programming, and probabilistic models to not only attribute an operation but also explain to the analyst why the system reaches its conclusions.

Belief Revision in Structured Probabilistic Argumentation

In real-world applications, knowledge bases consisting of all the information at hand for a specific domain, along with the current state of affairs, are bound to contain contradictory data coming from different sources, as well as data with varying degrees of uncertainty attached. Likewise, an important aspect of the effort associated with maintaining knowledge bases is deciding what information is no longer useful; pieces of information (such as intelligence reports) may be outdated, may come from sources that have recently been discovered to be of low quality, or abundant evidence may be available that contradicts them. In this paper, we propose a probabilistic structured argumentation framework that arises from the extension of Presumptive Defeasible Logic Programming (PreDeLP) with probabilistic models, and argue that this formalism is capable of addressing the basic issues of handling contradictory and uncertain data. Then, to address the last issue, we focus on the study of non-prioritized belief revision operations over probabilistic PreDeLP programs. We propose a set of rationality postulates -- based on well-known ones developed for classical knowledge bases -- that characterize how such operations should behave, and study a class of operators along with theoretical relationships with the proposed postulates, including a representation theorem stating the equivalence between this class and the class of operators characterized by the postulates.