Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Saptarshi Sengupta
Second Advisor
Navrati Saxena
Third Advisor
William Andreopoulos
Keywords
Time-Series Forecasting, TimesFM, causal tracing, Rankone- model-edit, model editing, projected gradient descent
Abstract
Time-series foundation models such as Google’s TimesFM 2.5 are increasingly used for forecasting in settings where downstream actions depend on the prediction. Like the deep networks they are built on, these models inherit adversarial weaknesses, yet almost all work on adversarial robustness has focused on language and vision. Defenses designed specifically for time series are scarce, and scarcer still are defenses that repair the model at the level of its weights rather than patching the inference stage. We take a two-stage approach. First, we use causal tracing to localize where in the network an adversarial perturbation on the input context becomes responsible for the corrupted forecast: the 1024-step context is split into four equal segments, the segment with the highest gradient sensitivity is chosen as the target, and the attack corrupts that segment. Second, we apply Rank-One Model Editing (ROME) to the single layer the trace identifies, recovering the clean forecast for the adversarial input through one rank-one update to an MLP weight matrix. On the UCI Electricity dataset, across four channels we observe post-edit MSE reductions of 96× to 860× under PGD, with specificity errors below 1.5×10−3. The same edit, located once, also repairs FGSM and BIM attacks across a range of budgets, indicating the correction is structural rather than tuned to a single attack.
Recommended Citation
Achar, Vaibhav, "Adversarial Robustness in Time-Series Foundation Models via Causal Tracing and Rank-One Model Editing" (2026). Master's Projects. 1768.
DOI: https://doi.org/10.31979/etd.3fmy-p4vv
https://scholarworks.sjsu.edu/etd_projects/1768
