TL;DR
This paper introduces FireCollaborator, a novel framework for proactive detection of Distributed Denial of Service (DDoS) attacks at the Internet Service Provider (ISP) level. By enabling Intrusion Prevention Systems (IPS) to collaborate and share compressed metrics like flow frequency and entropy, FireCollaborator aims to detect and mitigate attacks closer to their source, thereby saving network resources and protecting clients more effectively.
What Problem Does This Paper Address?
Distributed Denial of Service (DDoS) attacks pose a significant threat to the internet, overwhelming target systems with traffic from numerous compromised sources. Traditional detection methods often struggle because they are deployed too close to the victim, leading to an overload of traffic analysis. Furthermore, attackers often disguise malicious traffic as legitimate, making detection difficult. The paper addresses the challenge of detecting these attacks proactively and as early as possible in the network path, ideally at the ISP level, where a broader view of traffic is available and resources can be more efficiently managed. The authors highlight the need for a solution that benefits both ISPs (through a potential revenue stream) and their clients (by offloading the complexity of security management).
Methods
FireCollaborator employs a collaborative approach among Intrusion Prevention Systems (IPS) deployed within an ISP's network. The core idea is to leverage compressed metrics derived from network traffic, specifically the frequency and entropy of traffic matching specific routing rules.
- Subscription Protocol: Clients can subscribe to this proactive detection service from their ISP. A trusted server within the ISP manages subscriptions, issuing tokens to clients and informing participating IPSs (routers) about protected clients. This creates virtual rings of IPSs around each client.
- Compressed Metrics: Instead of analyzing every packet, IPSs focus on aggregated metrics.
- Frequency ($f_i$): The number of packets matching a specific rule $R_i$.
- Entropy ($H$): A measure of the distribution of traffic across different rules, calculated as $H = -\sum f_i \log f_i$. High entropy suggests a uniform distribution, while low entropy indicates a skewed distribution.
- Relative Entropy (Kullback-Leibler distance): Used to compare the current traffic distribution ($f$) with a historical profile ($f'$) to detect deviations.
- Detection Process:
- Selection Manager: Identifies rules with significant changes in frequency or entropy compared to a historical profile, using thresholds ($\gamma$, $\epsilon$) to filter out noise and focus on suspicious patterns.
- Score Manager: Assigns an "attack belief" score to selected rules based on combinations of high/low frequency and high/low entropy (categorized into cases 1, 2, and 3). These scores are aged and combined with scores from other IPSs using a belief combination function.
- Detection Manager: Rules with scores exceeding a threshold ($\tau$) are flagged as highly suspicious. The IPS then attempts to determine the actual packet rate and compares it to the client's capacity. If the rate exceeds capacity, an attack is confirmed. If not, the information is passed to neighboring IPSs for further analysis.
- Collaboration: IPSs communicate horizontally (within a ring) and vertically (towards the client) to share scores and build a more comprehensive picture of potential attacks. This collaborative filtering reduces false positives and allows for earlier detection.
Datasets
Not discussed in the paper. The evaluation was performed using simulated traffic, including both normal and embedded attack scenarios.
Key Insights and Ideas
- ISP-Level Detection is Key: Moving detection from the client's edge to the ISP level offers significant advantages in terms of visibility and resource management.
- Compressed Metrics are Powerful: Using frequency and entropy as proxies for traffic behavior is an efficient way to detect anomalies without deep packet inspection, making the system scalable.
- Collaboration is Crucial for Accuracy: Sharing detection scores and beliefs among multiple IPSs significantly improves the accuracy of attack detection and reduces false positives. The paper demonstrates that rings further from the victim (outer rings) can detect attacks earlier due to the aggregated traffic patterns.
- Proactive Service Model: The subscription-based model incentivizes ISPs to offer this security service, creating a win-win situation for both providers and customers.
- Multi-Level Filtering Reduces Overhead: The hierarchical detection process (selection, scoring, then detailed analysis) effectively filters out a large number of non-suspicious rules, reducing the computational burden on individual IPSs.
Key Results
The paper presents simulation results demonstrating the effectiveness of FireCollaborator:
- Improved Detection Rate: The system achieves a good detection rate, especially with a higher number of rings, indicating that collaboration is vital. For a 5-ring topology, a score threshold ($\tau$) of 0.7 provided a good balance between detection rate and false alerts.
- Reduced False Alerts: The multi-level approach, particularly the rule selection and scoring mechanisms, significantly reduces the number of false alerts compared to simpler detection methods.
- Early Detection: The system can detect attacks early in their duration, allowing for rapid response and mitigation, thereby reducing the impact on the client and the network.
- Efficiency: The collaborative approach and the use of compressed metrics lead to reduced communication overhead and less rule selection, making the system scalable.
- Aggressiveness Matters: More aggressive attacks are detected more successfully, with detection rates plateauing after a certain level of aggressiveness.
Limitations and Future Work
- Simulation-Based: The evaluation was conducted through simulations. The authors suggest using real network traces for future validation.
- Parameter Tuning: The paper notes that fixing the right parameters (like thresholds $\gamma$, $\epsilon$, $\tau$, and score factor) requires a learning stage and specific simulations.
- Security of IPS Communication: The authors acknowledge the need to address security concerns related to IPS communication itself, including the effects of compromised IPSs.
- Limited Participating Routers: The impact of a limited number of participating routers in the detection system is an area for future exploration.
- Real-time Performance: While the paper focuses on detection logic, further work could explore the real-time performance implications and optimizations.
Summary updated