Abstract
Marked Temporal Point Processes (MTPPs) are essential for modeling event sequences in continuous time across domains such as finance, healthcare, and social media. However, their reliance on chronological ordering makes them susceptible to unique adversarial vulnerabilities. This paper introduces PERMTPP, the first end-to-end differentiable framework for generating adversarial attacks against MTPPs. By combining Gumbel-Sinkhorn networks for soft permutations with attention-based temporal noise injection, PERMTPP effectively exposes deep vulnerabilities in state-of-the-art models that traditional gradient-based methods fail to capture.
Introduction
Adversarial attacks have been extensively explored in domains like computer vision and natural language processing. However, MTPPs—defined by sequences of event times and marks $(t_i, c_i)$—present unique hurdles. First, MTPP models (e.g., RNNs or Transformers) are highly sensitive to event ordering; a minor temporal shift that swaps two events can cause an "abrupt jump" in the model's hidden state, creating a non-smooth loss landscape. Second, finding the optimal adversarial reordering is a combinatorial problem with $n!$ complexity.
The PERMTPP Framework
To overcome these challenges, PERMTPP utilizes a two-stage differentiable pipeline that allows for optimization via gradient descent.
- Adversarial Permutation ($GS_\theta$): The framework employs a Gumbel-Sinkhorn network to relax discrete permutations into continuous operations. This allows the attacker to learn a "soft" permutation matrix that identifies the most disruptive event reorderings without exhaustive search.
- Temporal Noise Injection ($ATTN_\phi$): An attention-based network determines the precise additive noise ($\epsilon$) for each timestamp. To maintain chronological consistency with the chosen permutation, a hinge loss regularizer is used as a "differentiable bridge," ensuring the optimization landscape remains walkable for backpropagation.
The Perturbation Paradox
A critical finding of this work is the "Perturbation Paradox." Standard distance metrics can be misleading in the context of MTPPs.
As shown above, adding a large constant time shift to every event results in zero distance (if normalized), whereas a tiny shift that swaps two adjacent events can completely break a model’s internal representation and subsequent predictions. This highlights the importance of history-sensitive perturbations over simple magnitude-based noise.
Experimental Evaluation
The authors evaluated PERMTPP on four real-world datasets: Taobao (e-commerce), Twitter (social media), Electricity (energy logs), and Health (clinical records). In the Taobao dataset, PERMTPP significantly outperformed the strongest baseline (RTS-P), increasing the Mean Absolute Error (MAE) of time predictions by 36.5% and decreasing Mark Prediction Accuracy (MPA) by 16.8% in black-box scenarios. In white-box cases, mark prediction accuracy was reduced to below 10% for some datasets, demonstrating extreme vulnerability.
Conclusion and Future Work
PERMTPP establishes a new benchmark for evaluating the robustness of MTPP models. Beyond attacking, the authors demonstrated that adversarial training using PERMTPP-generated examples significantly enhances model resilience. Future research directions include extending the framework to generate entirely new synthetic marks and adapting the differentiable approach for privacy-preserving applications, such as differential privacy for event sequences.
Summary updated