Robustness analysis of location aware collaborative filtering for web service recommendation

Xiang Li, Min Gao, Wenge Rong, Junhao Wen, Qingyu Xiong2016 · DOI

Abstract

Web service recommendation systems increasingly rely on Location-Aware Collaborative Filtering (LACF) to improve Quality-of-Service (QoS) predictions by leveraging geographic proximity. However, the open nature of these systems makes them vulnerable to malicious data injection. This paper investigates the robustness of LACF against Location-Aware Shilling Attacks (LASA). By forging geographic identifiers like IP addresses, attackers can manipulate recommendation outcomes. Our analysis of three attack models—Random, Average, and Bandwagon—demonstrates that location-aware attacks, particularly the Average and Bandwagon models, significantly degrade recommendation accuracy compared to classical methods.

Introduction

Web service recommendation systems often rely on Collaborative Filtering (CF) to predict Quality-of-Service (QoS) metrics like response time. Because network performance is heavily tied to geography, Location-Aware Collaborative Filtering (LACF) has become a standard improvement. LACF prioritizes data from users in the same region or Autonomous System (AS) to provide more relevant predictions.

However, these systems are "open," meaning anyone can contribute data. This creates a surface for shilling attacks: malicious users inject fake profiles with biased QoS values to either promote their own services ("push attacks") or bad-mouth competitors ("nuke attacks"). While classical shilling attacks are well-documented, this research explores whether the addition of location data makes these attacks more dangerous and how attackers can exploit the geography-based logic of LACF.

Location-Aware Shilling Attacks (LASA)

The study introduces three new Location-Aware Shilling Attack (LASA) models designed to exploit the geographic prioritization of LACF systems. In these models, attackers not only inject biased QoS data but also forge their IP addresses to match the target user's location (AS and Country).

The structure of a malicious profile in a Location-Aware Shilling Attack, incorporating geographic identifiers (ASNu, Country ID) alongside QoS data.

The three models evaluated are:

  • Location-Aware Random (LAR): Attackers inject profiles with random QoS values while forging their location.
  • Location-Aware Average (LAA): Attackers use the mean QoS value for each service to appear as "typical" users, combined with forged location data.
  • Location-Aware Bandwagon (LAB): This exploits the popularity of certain services (Zipf’s law). Attackers target frequently invoked services to appear highly similar to a large number of real users.

Experimental Methodology

The researchers used the WSDREAM dataset, containing nearly 2 million service invocations from 339 users across 5,825 web services. To simulate real-world conditions, data density was reduced to 20%, with an 80/20 split for training and testing. The recommendation logic employed a user-based CF approach using the Pearson Correlation Coefficient (PCC) for similarity and K-Nearest Neighbors (KNN) for selection, with a preference for neighbors within the same AS or country.

Results and Analysis

The primary metric for success was the Average Prediction Shift, which measures the change in predicted response time after an attack.

Prediction shift as a function of attack size for the three location-aware attack models (LAA, LAB, LAR).

Key findings include:

  • Dominance of the Average Attack: The LAA model caused the most significant shifts, ranging from approximately 0.12 to 0.32 seconds. Given that most services in the dataset have response times under 1 second, this shift is substantial enough to move a service from "top-tier" to "unusable."
  • Synergy of Location and Popularity: The LAB attack significantly outperformed its classical counterpart. By combining geographic proximity with interest in popular services, attackers create profiles that the system's KNN logic finds difficult to exclude. The prediction shift for LAB reached up to 0.14 seconds, whereas the classical version barely impacted the system.
  • Vulnerability to Scaling: As shown in Figure 3, the prediction shift increases as the attack size (the percentage of fake profiles) grows from 3% to 15%, indicating a lack of natural resistance to scaled attacks.

Conclusion

Location-aware recommendation systems possess a geographic "backdoor." By forging an IP address to match a target user's Autonomous System, an attacker can force the engine to prioritize malicious profiles. While this study highlights these vulnerabilities, it also points to the need for future work in developing robust defense mechanisms and anomaly detection algorithms capable of identifying forged geographic data.

Summary updated

BibTeX
@misc{li2016robustness,
  author = {Xiang Li and Min Gao and Wenge Rong and Junhao Wen and Qingyu Xiong},
  title = {{Robustness analysis of location aware collaborative filtering for web service recommendation}},
  year = {2016},
  doi = {10.29268/iciot.2016.0003},
}
Figure