AI to face covert attacks in IoT and softwarized scenarios: challenges and opportunities

Abstract

Recently, the number of attacks aiming at breaching networked and softwarized environments has been growing exponentially. In particular, information hiding methods and covert attacks have been proven to be able to elude traditional detection systems and exfiltrate sensitive data without producing visible network flows or data exchanges. In this context, Artificial Intelligence techniques can play a key role in detecting these new emerging attacks, owing to their capability of quickly processing huge amounts of data without the necessity of expert intervention. In this work, we discuss the main challenges to face covert attacks in IoT and softwarized environments and we describe some preliminary results obtained by adopting Ital-IA 2023: 3rd National Conference on Artificial Intelligence, orga-

1. Introduction

The recent surge in the use of Artificial Intelligence (AI) is also supported by its adoption to face the growing number of cyber threats [ 1 ]. In fact, in recent years, attacks intensified both in terms of volume and complexity, for instance, by using techniques to elude detection or to conceal trafic flows exchanged with a remote controller [ 2 ]. Moreover, advanced persistent threats demonstrated their ability to bypass many security perimeters, also due to the use of multi-stage loading architectures or techniques to conceal attack routines. To this aim, AI demonstrated to represent an efective tool for the detection, reverse engineering, and forensics operations nized by CINI, May 29–31, 2023, Pisa, Italy ∗Corresponding author. operations [ 3 ].

However, an emerging aspect concerns the use of information hiding techniques to implement network covert channels. In this case, the attacker aims at eluding detection by not triggering classical defense systems based on trafic anomalies or exfiltrating sensitive data without producing visible flows. Owing to the efectiveness of the approach, threat actors are increasingly exploring new “carriers”, i.e., containers able to conceal secret and malicious data. In this vein, a very recent efort is devoted to understanding the feasibility of using the AI itself as a carrier for malicious data; see, e.g., [ 4 ] for a discussion on how information can be concealed in neurons or model parameters. To make things more complex, the increasing softwarization of networks and services accounts for an almost boundless attack surface that can be exploited to make future malware dificult to tame.

From this perspective, our paper showcases the most recent advances and research questions on the use of AI to mitigate malware leveraging covert communications or implementing (hidden) leakage attempts in containerized architectures. Specifically, it discusses the use of federated learning to eficiently deploy AI-based countermeasures in ubiquitous IoT scenarios, as well as the challenges characterizing micro-service architectures built around container technologies. In addition, to ofer a comprehensive discussion, we outline opportunities offered by graph generation to devise new efective solutions for detecting evolving threats.

Summing up, the contribution of this work is to shed

Covert Sender Legitimate Network Traffic Covert Channel

Network Packets with Hidden Information

Internet

Covert Receiver new light on the use of AI to face the emerging threat encoded by the covert endpoints with two distinguished endowed with covert attacks, especially when targeting TTL values. To reveal the presence of such covert chanrealistic scenarios based on IoT or container technolo- nels, we developed a detection mechanism based on AI. gies. Another contribution concerns the investigation In particular, we leveraged autoencoders because of their of “perspective” challenges given by the use of AI-based capability to also deal with attacks undocumented and frameworks. unknown a priori, as it happens when considering the

The rest of the paper is structured as follows. Section carrier used to create the covert channel. To perform 2 deals with AI to mitigate threats using covert channels the detection, we monitored “windows” of network packin IoT ecosystems, whereas Section 3 discusses oppor- ets to extract statistical metrics related to the TTL, e.g, tunities for improving the security of containerized en- the maximum, the minimum, or the average TTL values. vironments. Section 4 showcases opportunities arising Only legitimate trafic information has been given to the from graph generation, and 5 presents challenges and autoencoder to perform the training. By contrast, the opportunities of using AI for cybersecurity-related tasks. compromised trafic information has been used to evalLastly, Section 6 concludes the paper. uate the performance of the detection model. Results showcased the efectiveness of the AI-based approach, i.e., we obtained ∼91% and ∼94% for the accuracy and the 2. Covert Malware in IoT Scenarios precision, respectively. Despite the promising results, we extended the work evaluating an incremental learning Information hiding techniques are increasingly used by scheme based on an ensemble of autoencoders trained attackers to conceal malware in diferent carriers [ 2 ]. For on disjointed data chunks [ 6 ]. Figure 2 depicts the proexample, network covert channels, i.e., hidden commu- posed architecture. Compared to the results obtained by nications attempts nested within network trafic, can be using a single autoencoder, we obtained ∼95% both for used to secretly exfiltrate information or to elude well- the accuracy and the precision when using an ensemknown Intrusion Detection Systems (IDSes). Figure 1 ble of neural models. By using the incremental learning depicts a possible reference scenario. In particular, it scheme, the model can also be deployed on devices with shows the covert sender, e.g., a compromised node of an limited computational and storage resources, for instance IoT deployment, exchanging secret information via net- in home gateways or edge nodes. work features with the covert receiver, e.g., the Command Although the ensemble-based model allows for im& Control facility of the attacker. To do this, the sender proving the detection capabilities w.r.t. a single model, could directly conceal data within the header of a proto- it requires to set the ensemble size, i.e., the number of col or encode the information in the temporal evolution windows to consider in learning the model. In addition, of network packets belonging to a specific conversation. it will be trained only against the data available on a

Unfortunately, network covert channels are often ne- single edge node and the owners of the data could be not glected by standard security tools, thus revealing their inclined to share them. presence is mandatory to fully assess the security of a To overcome all these issues, we are interested in inmodern network. To this aim, we addressed the problem vestigating the usage of a federated learning approach of revealing hidden network communications targeting to address this task. In [ 7 ], we evaluated the benefits the IPv4 protocol in an IoT ecosystem [ 5 ]. We considered of this paradigm in a related (information-hiding) scethe exfiltration of data hidden in the Time To Live (TTL) nario in which malicious payloads are hidden within ifeld of a tampered IoT node, i.e., the bit “1” and “0” are

Data Stream various possible attacks, threat actors are increasingly exploiting techniques to let containers leak data beyond a well-defined execution perimeter [ 9 ]. This mechanism cDhautnaks Di-T D… Di-k D… Di irnestehmisbcleasset,hae schreaarteido nloocfaal rceosvoeurrtccehiasnanlteelr(esdeet,oSeecntcioodne2a): secret message within its temporal evolution, which can be observed outside the single container. For instance, a sender process can manipulate the amount of memory used within its container to alter the overall (host-level) Ensemble Model available memory. Then, in parallel, a receiving process running into another container can infer the message by Figure 2: Incremental Deep Ensemble model approach. inspecting such a global statistic, e.g., by reading an entry in the /proc/ filesystem for the case of Docker [ 9, 10 ].

Server The security implications of such an “imperfect” isolation UUppddaattee SLeorcvaelrMMooddelel Sedravtear taoreorscehveesrtarla.tFeiarsnt,acttoanctkaiannedrssycnacnheroxnchizaenmgea ninyfporromcaetsisoens to implement DDoS/slow-DoS attacks [ 10, 11 ]. Second, containers can leak information to recognize features of the underlying hardware/software infrastructure and support the reconnaissance stage at the basis of many Node1 Node2 Node… Noden complex attack chains [ 10 ].

Ldoactaal Ldoactaal Ldoactaal Ldoactaal appTroomacihti,geastpeescuiachllystiefajlothinytlayttuascekds, wAiIthistaooplrsoambilseintog gather precise information on the behavior of the overall Figure 3: Federated Learning Architecture. software architecture [ 12 ]. As an example, containers exchanging information in a secret manner could exhibit tight temporal correlations. This is due to the need of high-resolution icons that are commonly used in most altering a resource and “decode” changes close in time, popular mobile ecosystems, including, Android and iOS. mainly to avoid that other competing processes will disFigure 3 depicts the proposed approach. We assume to rupt the encoded secret [13]. Hence, AI can be used to have a centralized server acting the role of coordinator spot anomalies in the “wake-sleep” pattern of containers among nodes. The server contains a “weak” DNN de- or to enlighten possible correlations dificult to capture tector model trained on an initial dataset with a limited with the common sense, e.g., the distillation of signatures number of examples. In the early stage, this initial detec- in system calls used to access shared statistics exposed by tor is shared across n end nodes, which then fine-tune the kernel of the guest OS. Another important application their model against their own local data. To make the of AI concerns the creation of suitable whitelists/blackpredictor more robust and to find a global model in a lists. In fact, containers can also exhibit tight-coupled distributed manner, a subset of end nodes periodically interactions when part of the same service. The AI can be sends updates to the server containing the weights of then leveraged to whitelist containers expected to have each layer composing their local DNN. The coordinator overlapped or correlated timing behaviors and prevent aggregates the information received to build an ensemble too aggressive detection rules. Techniques like process model and again shared it with the peers. This process is mining can be used to define precise traces of the system iterated until a certain convergence criterion is reached. calls invoked by the software running within the various containers. This can allow for the early revealing of possible exfiltration attempts or feed an additional AI3. Container Security based framework to perform runtime detection of leaking attempts.

Containers are now the preferred choice for the creation Indeed, AI can also be employed to face other types of scalable frameworks able to take advantage of the of threats. For instance, the trafic exchanged by various micro-service paradigm. Specifically, their lightweight containers or softwarized architectures can be used to nature ofers many benefits compared to classical virtual detect exfiltration attempts, network-based attacks (e.g., machines, e.g., a smaller resource footprint or a reduced DoS/DDoS) or the presence of crypto-miners sending delay when deploying new services. Unfortunately, secu- data to a centralized master entity [14]. Besides, more rity requirements of containers are still not fully under- classical approaches based on the analysis of logs or the stood as for the case of virtual machines [ 8 ]. Among the (automatic) correlation of configurations and network trafic should be updated to handle the highly-mutable the graph structure require flexible architectures able to and dynamic nature of containerized services. Despite handle such modifications. We plan to devise a deepthe considered threat, micro-service architectures imple- learning-based approach aiming at predicting graph evomented via containers present many research challenges lution by considering stepwise changes (that can afect and open issues, which are very stimulating. On one both, node and edge set). Moreover, to model long-term hand, the need of gathering, processing and storing traf- evolutions, we are interested in defining an architecture ifc at “wire speed” without degrading the overall Quality invariant to the network dimension. Learning the evoluof Experience for feeding the AI poses many technical tion of dynamic graphs by extracting their temporal and challenges. On the other hand, the presence of sensitive structural characteristics could be useful to identify polydata and the need of conforming to constraints like the morphic cyber attacks, such as polymorphic malware [21] General Data Protection Regulation (GDPR) impose to or those leveraging some form of obfuscation. This can minimize the disclosure of personal or de-anonymizable also help to assess the emerging wave of steganographic bits. The GDPR also requires to consider both the phys- malware [22]. In essence, this class of threats leverages ical and legal boundaries where data is stored and pro- information hiding techniques to conceal malicious ascessed. In this case, softwarized networks and infras- sets (e.g., configuration files or additional attack stages) tructures are expected to greatly benefit from federated in various software artifacts, such as images, executables approaches, especially owing to the edge flavor of many or metadata. Unfortunately, being very threat-specific, future scenarios such as those based on 5G (see, e.g., [15] generalizing their detection via standard mechanisms or and the references therein). signature-based approaches is a challenging task. In this

Lastly, container security will also benefit from threat- vein, graph generation could lead to a more abstract and specific hardening and configuration mechanisms. Also unified framework to early detect malicious software tryin this case, AI represents a stimulating asset. For in- ing to reduce its footprint to remain unnoticed for long stance, part of the software development life cycle of time frames. containerized applications [16] can take advantage of AI, especially to evaluate configurations, impose hardening constraints or match security issues against known CVEs 5. Future and Main Challenges or taxonomies. For the specific case of covert attacks, a promising use case for AI concerns the identification of resources that can be (ab)used to encode secret information and characterized by “loose” isolation properties. As a paradigmatic example, a suitable resource matrix [17] can be computed by inspecting containers and software (e.g., used libraries, code patterns, or execution privileges) to provide automatic configuration policies preventing leaking behaviors. This represents a major research challenge, especially to avoid that the container engine will terminate or impede the execution of “flagged” containers in a too aggressive manner, thus impairing the quality of the overall service.

To efectively deploy AI for taming covert and hidden

threats, various challenges and refinements have to be addressed. First, the obtained data should model as faithfully as possible real scenarios dealing with cybersecurity.

This is in general complex, but it becomes exceptionally hard when considering information-hiding-capable threats. In fact, the lack of publicly-available datasets capturing the presence of malware using covert channels or other elusive mechanisms is a well-known problem [ 2, 22 ]. Moreover, in publicly-available datasets samples that represent cyber attacks are rare. Hence, the class imbalance can dramatically afect the performances of the detection models. Therefore, it is crucial to devise approaches able to handle this skewness, e.g., by exploiting 4. Graph Generation for Detection generative models able to produce realistic anomalies. Again, real data are dificult to retrieve, and, moreover, In real-world scenarios, such as social networks, biology, they are usually afected by noise. In this respect, unand recommender systems, complex relations among the labelled data can lead the systems to make errors if not entities of graphs can hardly be modeled as flat tabu- properly handled as anomalies could be labeled as “norlar data. In addition, many current AI-based solutions mal”, and, vice versa. In addition, due to data scarcity assume that the underlying graph is static, however real- and skewness, the systems could erroneously classify inworld networks are dynamic as both their topology and frequent legit behaviors (due to data shifts) as anomalous dimension tend to change over time. Dynamic graphs ones. [18] are typically adopted in several application domains, Moreover, some of the techniques briefly discussed including cybersecurity with diferent purposes such as in this work [ 5, 6, 7 ] are threat-dependent and dificult malware [19] and intrusion detection [20]. Dynamic evo- to generalize. For instance, AI-based models are able lution is dificult to model due to the dynamic nature to detect hidden communication attempts only in fields of the underlying process, where continuous changes in of the IPv4 protocol sharing similar functionalities [ 6 ]. Therefore, a more general approach should be considered to devise flexible and robust models. A possible idea is to use more abstract metrics typical of at least a class of hiding mechanisms. Then, multiple AI pipelines can be deployed to eficiently detect a family of attacks, e.g., methods targeting IPv4 or IPv6 conversations [23]. In addition, in unsupervised settings, it is dificult to define boundaries between normal and anomalous behaviors.

Again, this is especially hard when considering covert attacks, which are stealthy by-design. Similarly, challenges have to be faced when considering container security.

On one hand, the overall security model still needs to be fully understood. On the other hand, the complexity of the various software layers requires proper modeling to eficiently provide data to the AI.

Lastly, as cyber-attacks can be represented via dynamic graphs, it is crucial to define strategies able to detect threats in evolving networks. Modeling and predicting the evolution of dynamic graphs is a challenging task due to their evolving nature. State-of-the-art systems lack lfexibility, and, in this respect, models that guarantee invariance w.r.t the input size should be devised.

6. Conclusions

This paper discussed the use of AI to address emerging

challenges, i.e., threats endowed with mechanisms able to covertly leak data in networked and softwarized environments. As discussed, each scenario requires facing diferent challenges. For instance, when dealing with IoT ecosystems, federated learning could be considered a main “technology enabler”, especially for distributing computation and guaranteeing that sensitive data remain confined at the border of the network. Instead, containerized architecture could benefit from AI to find correlations and patterns in the complex interplay of software components.

Unfortunately, the AI could also be exploited to hide data, thus spawning new threats. To this aim, graph generation should be carefully considered as it can help in ifnding a convenient representation of operations (both at the abstract level or in terms of system calls) to support the detection of advanced ofensive schemes. Accordingly, part of our future research will address the aforementioned topics.

Acknowledgments

This work was partially supported by project SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU.

[13] C. Marforio, H. Ritzdorf, A. Francillon, S. Capkun, representation learning and generative modeling, Analysis of the communication between colluding 2022. doi:10.48550/ARXIV.2208.12126. applications on modern smartphones, in: Proceed- [19] B. Anderson, D. Quist, J. Neil, C. Storlie, T. Lane, ings of the 28th Annual Computer Security Appli- Graph-based malware detection using dynamic cations Conference, 2012, pp. 51–60. analysis, Journal in computer Virology 7 (2011) [14] J. A. Perez-Diaz, I. A. Valdovinos, K.-K. R. Choo, 247–258.

D. Zhu, A flexible sdn-based architecture for [20] G. Duan, H. Lv, H. Wang, G. Feng, Application identifying and mitigating low-rate ddos attacks of a dynamic line graph neural network for intruusing machine learning, IEEE Access 8 (2020) sion detection with semisupervised learning, IEEE 155859–155872. Transactions on Information Forensics and Security [15] V. Rey, P. M. S. Sánchez, A. H. Celdrán, G. Bovet, 18 (2023) 699–714.

Federated learning for malware detection in iot de- [21] E. Avllazagaj, Z. Zhu, L. Bilge, D. Balzarotti, T. Duvices, Computer Networks 204 (2022) 108693. mitras, When malware changed its mind: An em[16] T. Rangnau, R. v. Buijtenen, F. Fransen, F. Turk- pirical study of variable program behaviors in the men, Continuous security testing: A case study real world., in: USENIX Security Symposium, 2021, on integrating dynamic security testing tools in pp. 3487–3504. ci/cd pipelines, in: 2020 IEEE 24th International En- [22] L. Caviglione, W. Mazurczyk, Never mind the malterprise Distributed Object Computing Conference ware, here’s the stegomalware, IEEE Security & (EDOC), IEEE, 2020, pp. 145–154. Privacy 20 (2022) 101–106. [17] R. A. Kemmerer, Shared resource matrix methodol- [23] W. Mazurczyk, K. Powójski, L. Caviglione, IPv6 ogy: An approach to identifying storage and timing covert channels in the wild, in: Proceedings of the channels, ACM Transactions on Computer Systems third central european cybersecurity conference, 1 (1983) 256–277. 2019, pp. 1–6. [18] S. Gupta, S. Bedathur, A survey on temporal graph

References

[1] M. J. H. Faruk , H. Shahriar , M. Valero , F. L. Barsha , S. Sobhan , M. A. Khan , M. Whitman , A. Cuzzocrea , D. Lo , A. Rahman , et al., Malware detection and prevention using artificial intelligence techniques , in: 2021 IEEE International Conference on Big Data (Big Data) , IEEE, 2021 , pp. 5369 - 5377 .

[2] W. Mazurczyk , L. Caviglione , Information Hiding as a Challenge for Malware Detection , IEEE Security & Privacy 2 ( 2015 ) 89 - 93 .

[3] D. Ucci , L. Aniello , R. Baldoni , Survey of machine learning techniques for malware analysis , Computers & Security 81 ( 2019 ) 123 - 147 .

[4] T. Liu , Z. Liu , Q. Liu , W. Wen , W. Xu , M. Li , Stegonet: Turn deep neural network into a stegomalware , in: Annual Computer Security Applications Conference, 2020 , pp. 928 - 938 .

[5] M. Guarascio , M. Zuppelli , N. Cassavia , G. Manco , L. Caviglione, Detection of Network Covert Channels in IoT Ecosystems Using Machine Learning , in: Proc. of The Italian Conference on CyberSecurity , volume 3260 of CEUR Workshop Proceedings , 2022 , pp. 102 - 113 .

[6] N. Cassavia , L. Caviglione , M. Guarascio , A. Liguori , M. Zuppelli , Ensembling Sparse Autoencoders for Network Covert Channel Detection in IoT Ecosystems , in: Foundations of Intelligent Systems: 26th International Symposium , Springer, 2022 , pp. 209 - 218 .

[7] N. Cassavia , L. Caviglione , M. Guarascio , A. Liguori , G. Surace, M. Zuppelli , Federated learning for the eficient detection of steganographic threats hidden in image icons , in: Pervasive Knowledge and Collective Intelligence on Web and Social Media , Springer Nature Switzerland, Cham, 2023 , pp. 83 - 95 .

[8] S. Sultan , I. Ahmad, T. Dimitriou, Container security: Issues, challenges, and the road ahead , IEEE Access 7 ( 2019 ) 52976 - 52996 .

[9] Y. Luo , W. Luo , X. Sun , Q. Shen , A. Ruan , Z. Wu , Whispers between the containers: High-capacity covert channel attacks in docker , in: Trustcom/BigDataSE/ISPA, IEEE, 2016 , pp. 630 - 637 .

[10] X. Gao , B. Steenkamer , Z. Gu , M. Kayaalp , D. Pendarakis , H. Wang , A study on the security implications of information leakages in container clouds , IEEE Transactions on Dependable and Secure Computing 18 ( 2018 ) 174 - 191 .

[11] I. Vaccari , M. Aiello , E. Cambiaso, Slowtt: A slow denial of service against iot networks , Information 11 ( 2020 ) 452 .

[12] F. Al-Doghman , N. Moustafa , I. Khalil , Z. Tari , A. Zomaya , Ai-enabled secure microservices in edge computing: Opportunities and challenges , IEEE Transactions on Services Computing ( 2022 ).