The Threat of Intelligent Attackers Using Deep Learning: The Backoff Attack Case

The Threat of Intelligent Attackers Using Deep Learning: The Backoff Attack Case

Juan Parras, Santiago Zazo
DOI: 10.4018/978-1-7998-5068-7.ch006
(Individual Chapters)
No Current Special Offers


The significant increase in the number of interconnected devices has brought new services and applications, as well as new network vulnerabilities. The increasing hardware capacities of these devices and the developments in the artificial intelligence field mean that new and complex attack methods are being developed. This chapter focuses on the backoff attack in a wireless network using CSMA/CA multiple access, and it shows that an intelligent attacker, making use of control theory, can successfully exploit a sequential probability ratio test-based defense mechanism. Also, recent developments in the deep reinforcement learning field allows that attackers that do not have full knowledge of the defense mechanism are able to successfully learn to attack it. Thus, this chapter illustrates by means of the backoff attack, the possibilities that the recent advances in the artificial intelligence field bring to intelligent attackers, and highlights the importance of researching in intelligent defense methods able to cope with such attackers.
Chapter Preview


This Chapter studies a threat that the current advances in Deep Learning pose to the security of Wireless Sensor Networks (WSNs). Namely, the Chapter focuses on the fact that Deep Reinforcement Learning (Deep RL) tools can be used to exploit a possibly unknown defense mechanism simply by interacting with it. The remarkable advances and proliferation in wireless networks in the last years have brought a significant interest in the security and threats to WSNs: they can be the target of many attacks due to the limited capabilities of the sensors, as some recent surveys show (Fragkiadakis, Tragos, & Askoxylakis, 2013) (Zhang, et al., 2015). One of these attacks is the backoff attack, which affects to the Medium Access Control (MAC) layer when a CSMA/CA (carrier-sense medium access with collision avoidance) scheme is used to regulate the access to the medium. The backoff mechanism minimizes the risk of collision, i.e., that two or more sensors transmit simultaneously, by deferring transmissions during a certain random time period: the backoff window. In a backoff attack, a sensor uses a lower backoff window than the rest of the sensors, thus obtaining a higher throughput at expense of the other sensors (Bayraktaroglu, et al., 2013).

In order to clearly study and explain the intelligent attacks that Deep RL tools bring, this Chapter focuses only in the backoff mechanism, as backoff attacks are a real threat to WSNs. Firstly, because network adapters are highly programmable (Cagalj, Ganeriwal, Aad, & Hubaux, 2005), thus allowing sensors to modify their backoff parameters. And secondly, because many MAC layer protocols proposed for WSNs make use of CSMA as medium access mechanism, for instance, SMAC (Ye, Heidemann, & Estrin, 2004), WiseMAC (Enz, El-Hoiydi, Decotignie, & Peiris, 2004), TMAC (Van Dam & Langendoen, 2003) and DSMAC (Lin, Qiao, & Wang, 2004). Actually, two surveys on MAC layer protocols, (Demirkol, Ersoy, & Alagoz, 2006) and (Yadav, Varma, Malaviya, & others, 2009), show that CSMA is the most common access mechanism in contention based MAC protocols.

Our main contribution consists in highlighting the effect that Deep RL based attackers can pose to current WSN defense mechanisms: such an attacker needs not knowing the defense mechanism used, nor its parameters, as it may learn to exploit it simply by interacting with it. Thus, we show that Deep RL based attackers are flexible due to their learning capabilities and pose a significant threat to many current WSN defense mechanisms.

Hence, the main objectives of the Chapter are: (1) introducing the backoff attack and its effects on the network throughput, (2) show different ways in which a game theory based defense mechanism may cope with such attack, (3) introduce a control theory formulation that allows obtaining an optimal attacker control law to exploit the defense mechanism proposed when the attacker knows all the parameters of the defense mechanism, and (4) show how Deep Reinforcement Learning tools can be used to successfully exploit the defense mechanism when the attacker does not know all the parameters of the defense mechanism.

Regarding the Chapter organization, in the Background Section we introduce the CSMA/CA mechanism, as well as how a backoff attack affects the network throughput, and the main control framework that will be used along this work: Markov Decision Processes. Then, we dedicate a Section to study defense mechanisms against the backoff attack: where we focus on sequential tests as an advanced defense mechanism. The next Section presents an optimal attack against this defense mechanism in the case that the defense mechanism is known, and also presents a Deep RL attack that is successful when the defense mechanism is unknown. The final Section summarizes our results and proposes some future lines of interest.

Key Terms in this Chapter

Hypothesis Test: Mathematical tool used to decide whether a certain data follows a certain distribution or not.

Good Sensor: Sensor in a CSMA/CA WSN that always respects the binary exponential backoff procedure.

Sequential Probability Ratio Test: Hypothesis test of variable sample size in which the test statistic is updated as new samples arrive and makes a decision once it has collected enough information.

Reinforcement Learning: Brach of the Artificial Intelligence field devoted to obtaining optimal control sequences for agents only by interacting with a concrete dynamical system.

Attacking Sensor: Sensor in a CSMA/CA WSN that may choose between respecting the binary exponential backoff procedure or using a uniform backoff procedure that provides the sensor with an advantage in terms of throughput.

Markov Decision Process: Mathematical framework used to model dynamical systems and obtain their optimal control policies.

Deep Neural Network: Graph structure which provides as output a nonlinear combination of its inputs. It has been proven to be a universal function approximator.

Complete Chapter List

Search this Book: