Building effective defensive deception requires understanding the complexity of the target environment, the nature and processes of attacks, as well as strengths and weaknesses of existing countermeasures. This is often not studied and detailed and only remain general in existing studies. For instance, APT attacks are complex and performed in multi-stage and the system consists of cyber and physical layers. This thrust therefore provisions game-theoretic and machine learning defensive deception techniques for real systems. The following specifies activities in this thrust.

This activity consists of investigating the parameters related to the network architecture (such as the points of vulnerabilities, network layers) when it comes to protection.

In this task, APT techniques are dissected while considering the complex nature of deployed networks relying on attack representations (attack graphs, networks), inter-relations of domains and different episodes of the attacks.

This task consists of formalizing aspects characterizing APT defenses such as intrusion detection systems, incident responses and firewalls, in designing static and dynamic games for CPS.

This activity consists of formalizing aspects characterizing APT defensive deception in static and dynamic games.

In this thrust, computational and artificial intelligence artifacts are coupled to automate the evolution of the defensive deception process and to build the subsequent memory that a defender can rely on. This work package is intended to leverage ML-based techniques resilient to adversarial attacks and able to capture deceptive objects, real objects and malicious objects profiles in the system. The contribution of this thrust is five-folds: (i) to couple computational and artificial intelligence techniques to create and attach temporal and robust/reliable profiles of any component involved in the interaction between the attacker and the defender; (ii) to build a model relying on previous profiles that virtually represents the whole deception games including all the interactions between the players and inter-layered flows; (iii) to investigate transfer learning and other techniques that can optimize the learning processes during the deception so that the defender can quickly observe and anticipate (iv) to model possible adversarial attacks exploited to deteriorate the performance of classifiers, and to design adequate mitigation approaches (v) to consider the real-time aspects of the deceptive process to build evolutive intelligence that can support guide the defender on the actions to perform based on the intents of the attacker. The following gives more details on each research sub- thrust.

The possible paths of attack across layers will be represented using a cyber-attack graph that is generated based on the vulnerabilities associated with each node of the network. The attack graphs are a direct graph made up of nodes that are pre- or post-conditions of an exploit, and edges referring to consequences of having a pre-condition that enables an exploit post-condition. Attack graph form knowledge of the attacker. The defender’s goal that is to take deceptive actions to prevent the attacker from taking control over the network resources can be designed as deceptive graph guided by the attack graph structure. Knowledge is built based on attack graphs that have the possible paths from the entry point (cyber, physical, node or layer) to the target point. The deceptive graph is accordingly built with adequate attributes based on placement of decoys. During the whole process, these graphs are dynamically fed and recorded as conditional rules to guide movement between nodes. The concept drift will also be monitored as well as mitigation methods such as Adaptive WINdowing (ADWIN) (Bifet & Gavalda 2007) will be experimented to assure temporality and robustness. SDN techniques with their capability to abstract networks in programmable software to control security’s aspects, are expected to be to generate large amounts of traffic flow data, and to train ML models to identify real to fake decoy objects. Mechanisms are proposed to profile machine learning adversarial attacks based on the attacker strategies. In literature, authors do not consider that it is possible for an attacker to plan evade from ML detectors. This aspect should be considered when designing the games. After profiling, this activity should minimize ML attack effects on the delivery performance.

In this activity, generative models such as Defense- GAN relying on Generative Adversarial Network (GAN) are studied to protect the different classifiers against adversarial attacks. But also, these techniques will be explored to intentionally generate fake images of the systems, to augment to concealment to the attacker (Lopes Antunes and Llopis Sanchez 2023). This involves the consideration of the vulnerable parameters of machine learning possibly exploited by the attacker, to build the deception games.

In this activity, transfer learning will be explored to avoid re-training that costs resources. But the constraint here would be to find a model that optimally fits with the deceptive environment and the type of attacks. Incremental and online learning would be investigated to look for schemes enabling addition of current instance to the previous knowledge and even new detected class of attack, instead of re- generating the knowledge from scratch. The techniques of feature relevance would also be useful to compress the information necessary to build the deception and attack intelligence.

In this activity, deception games based on RL and DRL techniques such as Markov Decision processes (MDP) and Q-learning will be designed to find an optimal policy of deployment for deceptive resources such as honeypots in different domains or layers in the CPS (Li et al. 2022). This orientation has three main advantages: exploiting RL rewards to formulate players’ utility functions, designing an evolutive learning without prior datasets and hybridizing game with ML parameters. RL and DRL require manually defined of behaviours through reward functions that constitutes learning from the environment. This process takes lot of time and therefore not appropriate to define all possible set of rules and rewards for environments necessitating a high degree of flexibility and adaptability. For systems in domains of military domains Imitation Learning (IL) techniques (Ogenyi et al. 2019; Zare et al. 2023) might be exploited to provide demonstrations, thus eliminating the development of explicit reward functions. IL could be leveraged to accelerate the learning during deception and to design the utility functions based by imitating an expert’s behaviour through the provision of demonstrations.