RESEARCH INTERESTS
- Inference/Communication Networks
- Security & Privacy
- Behavioral Systems & Human Decision Making
- Network Economics
PROJECTS
Security in Inference Networks
Inference networks consist of spatially-distributed sensing agents that collect and transmit observations to a central node called the fusion center (FC), so that a global inference is made regarding the phenomenon-of-interest. Distributed inference networks are those where the sensors compress their observations in order to have band-limited transmissions. Since these sensing agents are simple and inexpensive, a distributed inference network is particularly vulnerable to security threats. So, we have investigated three types of security attacks, namely Eavesdropping, Byzantine (data-falsification) and Jamming attacks, within the context of inference networks.
In the context of eavesdropping attacks in detection networks, we have proposed optimal stochastic ciphers and designed near-optimal transmit-diversity strategies under constrained Eve’s performance. We have also designed optimal sensor decision rules under both differential and constrained secrecy frameworks. In the context of Byzantine attacks in distributed detection/estimation networks, we have shown tremendous improvements in network-security when the sensors employ finer quantization, and found optimal attacker strategies under resource-constrained settings. From the network’s perspective, we have also proposed mitigation schemes based on anomaly detection and noise-enhanced detection. In the context of jamming attacks, we have modeled a zero-sum game between a detection network and a jammer and investigated equilibrium solutions when sensors employ decode- forward/amplify-forward strategies, and, when the FC employs receiver-combining on the copies of received signals obtained using multiple antennas.
Spectrum Allocation in Cognitive-Radio Networks
Cognitive radio (CR) networks have been proposed to address the problem of spectrum scarcity and facilitate new wireless applications by letting the CRs use already licensed spectrum without interfering with the primary users (PUs). In the past, spectrum auctions have been designed under the assumption that PUs themselves moderate the allocation process. In settings where PUs cannot carry out the allocation process, they outsource this task to a moderator who may not have complete knowledge about the dynamic spectrum activity at the PUs.
Therefore, we have investigated the design of optimal auctions and bilateral trading mechanisms for CR networks, when the moderator has uncertainty regarding the availability of spectrum. One way to reduce moderator’s uncertainty is to let CRs collect observations regarding PU activity and perform cooperative spectrum sensing at the moderator. Such a design leads us to the notion of multidimensional truthfulness in mechanism design, since CRs can possibly falsify both spectrum decisions (compressed observations) and valuations before revealing them to the moderator. We have designed a novel sealed-bid auction that maximizes the moderator’s utility while guaranteeing truthful revelation of both sensing decisions and spectrum valuations. In the presence of multiple PUs, we have proposed a bilateral trading framework where the moderator allocates the PU spectra iteratively to the CRs depending on sensing data and spectrum valuations, such that the revenue of the moderator is improved at every time iteration.
Information Dispersal Games in Communication Networks
Information dispersal algorithms (IDAs) disperse a given source message across multiple paths in a communication/storage network, in an attempt to send the message to a destination node in a secure and reliable manner. One of the well-known IDA is the Rabin’s IDA [3] which simultaneously addresses secrecy and fault-tolerance by encoding a data le and decomposing it into unrecognizable data-packets before transmitting or storing them in a network over multiple paths. In our work, we have modeled Rabin’s framework as a zero-sum game between the source and the attacker, where the source picks an encoding scheme and disperses the decomposed codeword across multiple paths, while attacker chooses which nodes to compromise in order to simultaneously decode the original message and falsify as many encoded symbols as possible. We have numerically evaluated the mixed-strategy Nash-Equilibrium by decoupling the bilinear game into two linear programs which have a primal-dual relationship.
Strategic Information Transmission between Prospect Theoretic Agents
Communication systems have been studied extensively for several decades when the motives of transmitter and the receiver are aligned with each other. However, such a synergy does not always exist in the real world, as observed in the case of markets and media. While Crawford and Sobel studied Nash equilibria for strategic information transmission (SIT) games, Akyol et al. have investigated Stackelberg equilibria in SIT games (with transmitter as the leader, and receiver as the follower), both in the presence of a deceptive transmitter. On the other hand, Kamenica and Gentzkow have modeled this framework as a Bayesian persuasion mechanism, where they have analyzed necessary and sufficient conditions for the existence of a mechanism that strictly benefits the transmitter.
In this project, we analyze the effects of human biases on equilibrium strategies in a Stackelberg game, when both the transmitter and the receiver are modeled as PT agents. In the presence of Gaussian source signal and a Gaussian test channel, we have shown that weight functions have no impact on the equilibrium strategies. In the presence of exponential source signals and an exponential test channel, we found that the weight function plays a role in the construction of equilibrium strategies at both the transmitter and the receiver. Currently, we are also looking into the effects of human biases on a deceptive transmitter when the source and message signals are multi-dimensional.
Statistical Inference by Prospect Theoretic Agents
With the advent of crowd-sensing systems, human agents have become an integral part of the sensing infrastructure within inference networks. Traditional inference networks are designed for rational sensing agents whose decisions are modeled to optimize an objective function based on expected utility theory (EUT). However, there is extensive evidence in the psychology literature on how humans exhibit deviating behaviors that cannot be justified by EUT. To our rescue, Kahneman and Tversky had successfully extended EUT using a descriptive model called prospect theory (PT) to account for several cognitive behaviors that cannot be explained using EUT. Therefore, we are investigating optimal inference rules employed by prospect-theoretic agents in the context of binary hypothesis testing and parameter estimation, to study the impact of cognitive biases on their inference performance.
Learning Commuter Preferences across Multiple Attributes from Revealed Preferences
Traditionally, human preferences are modeled using discrete choice theory, which inherently makes strong assumptions such as transitivity and substitutability, which are shown to be violated by human agents in the psychology literature. Therefore, we study the possibility of modeling and identifying agents’ preferences across multiple attributes from choices revealed over a sequence of experiments in order to preserve some unique features of pro-sociality and informational attributes in human decision making. We are currently investigating an active data acquisition mechanism where the controller sends a sequence of strategic signals to evaluate the agent’s preferences across attributes. We have proposed a learning algorithm that estimates the agent’s preferences across multiple attributes from the data collected using the above strategic data acquisition mechanism.
THESIS & DISSERTATIONS
PhD Dissertation
Title: On the Design and Analysis of Secure Inference Networks
Advisor: Prof. Pramod K. Varshney
Department of Electrical Engineering and Computer Science, Syracuse University
MS Thesis
Title: Secure Distributed Detection in Wireless Sensor Networks via Encryption of Sensor Decisions
Advisor: Prof. Morteza Naraghi-Pour
Department of Electrical and Computer Engineering, Louisiana State University
