Abstract: In this paper we describe an algorithm for predicting the websites at risk in a long range hacking activity, while jointly inferring the provenance and evolution of vulnerabilities on websites over continuous time. Specifically, we use hazard regression with a time-varying additive hazard function parameterized in a generalized linear form. The activation coefficients on each feature are continuous-time functions constrained with total variation penalty inspired by hacking campaigns. We show that the optimal solution is a 0th order spline with a finite number of adaptively chosen knots, and can be solved efficiently. Experiments on real data show that our method significantly outperforms classic methods while providing meaningful interpretability.
Introduction
Predicting website compromises is a dynamic challenge. Traditional survival analysis often employs the Cox Proportional Hazards model, which assumes that the risk contribution of a specific feature—such as a specific software version—remains static over time. However, cyberattacks are inherently episodic. Hacking campaigns typically follow a lifecycle: an exploit is discovered, weaponized into automated kits, and eventually mitigated as patches are deployed or targets are exhausted.
Existing models struggle to capture these discrete "jumps" in risk levels across high-dimensional feature spaces. This research introduces a model capable of pinpointing exactly when a technology becomes a liability and when that danger subsides by jointly inferring the provenance and evolution of vulnerabilities.
Methodology
The proposed approach utilizes an Additive Hazard Function where the total hazard rate is the sum of contributions from various features. The core innovation is the treatment of feature coefficients $w$ as time-varying functions $w(t)$.
Total Variation Penalty
To capture the "bursty" nature of hacking, a Total Variation (TV) penalty is applied to these functions. In optimization, TV penalties encourage piecewise constant solutions. This aligns with domain knowledge: the risk associated with a vulnerability stays flat until a campaign begins (a "jump") and potentially returns to zero once the campaign ends.
Optimization and Efficiency
The authors prove a Representer Theorem showing that the optimal solution is a 0th order spline (a step function) with a finite number of knots. This theoretical result allows the complex functional optimization to be reformulated as a tractable, finite-dimensional fused lasso problem. This ensures computational efficiency even when dealing with hundreds of thousands of features.
Experimental Setup
The model was evaluated on a large-scale real-world dataset comprising:
- Hacked Sites: ~50,000 entries from blacklists (2011–2013), including phishing and search redirection.
- Benign Sites: ~337,000 sites sampled from
.comzone files, treated as right-censored data. - Features: Over 159,000 features extracted from HTML metadata (e.g., CMS versions).
Results and Discussion
Predictive Performance
The proposed "$\ell_1$ + non-monotone" model significantly outperforms the standard Cox Proportional Hazards model. By allowing the hazard rate to decrease (non-monotone), the model more accurately reflects the lifecycle of a hacking campaign.
Interpretability and Latent Hazard Estimation
A key advantage of this framework is its ability to discover the "provenance" of attacks. By examining the inferred hazard functions for specific features, researchers can visualize the timeline of vulnerabilities.
As shown in the hazard rate plots, the model identifies distinct periods of high risk. For instance, WordPress 3.5.1 shows a small-scale "proof of concept" spike followed by a massive campaign later in the year. Traditional models would average these risks, obscuring the actual threat landscape.
Conclusion
This work demonstrates that modeling hacking as a discrete, time-varying process significantly improves both predictive accuracy and interpretability. The use of a TV penalty effectively captures the episodic nature of cyber threats. Future work may explore low-rank structures to handle feature correlations and the integration of deep learning for non-linear feature interactions.
Summary updated