Decentralized Detection in Sensor Networks with Noisy Communication Links ⋆ Gianluigi Ferrari and Roberto Pagliari University of Parma, Parco Area delle Scienze 181A - 43100 Parma, Italy gianluigi.ferrari@unipr.it,pagliari@tlc.unipr.it WWW home page: http://www.tlc.unipr.it/ferrari Abstract. This paper presents a general approach to distributed de- tection with multiple sensors in network scenarios with noisy communi- cation links between the sensors and the fusion center (or access point, AP). The sensors are independent and observe a common phenomenon. While in most of the literature the performance metrics usually consid- ered are missed detection and false alarm probabilities, in this paper we follow a Bayesian approach for the evaluation of the probability of de- cision error at the AP. We first derive an optimized fusion rule at the AP in a scenario with ideal communication links. Then, we consider the presence of noisy links and model them as binary symmetric channels (BSCs). This assumption leads to a simple, yet meaningful, performance analysis. Under this assumption, we show, both analytically and through simulations, that if the noise intensity is above a critical level (i.e., the cross-over probability of the BSC is above a critical value), the lowest probability of decision error at the AP is obtained if the AP selectively discards the information transmitted by the sensors with noisy links. Key words: Decentralized detection, sensor networks, noisy communi- cation links, multiple observations, cross-layer design. 1 Introduction Distributed detection has been an active research field for a long time [19]. In particular, several approaches have been proposed to study this problem, in the realms of information theory [10], target recognition [16, 17], and several other areas. The increasing interest, over the last decade, for sensor networks, has spurred a significant research activity on distributed detection techniques in this context [21, 4, 7, 9]. In recent years, wireless sensor networks are becoming more common, as, for example, in terrain monitoring applications [18]. In a wireless communica- tion scenario, links between sensors and the access point (AP) are likely to be faded [15, 6]. In this case, most of the results proposed in the literature are not ⋆ This work was supported in part by Ministero dell’Istruzione, Universit`a e Ricerca (MIUR), Italy, under the PRIN project “CRIMSON” (Cooperative Remote Inter- connected Measurement Systems Over Networks).