Beyond Immediate Activation: Temporally Decoupled Backdoor Attacks on Time Series Forecasting
Zhixin Liu, Xuanlin Liu, Sihan Xu, Yaqiong Qiao, Ying Zhang, Xiangrui Cai · 2026
Read Beyond Immediate Activation: Temporally Decoupled Backdoor Attacks on Time Series Forecasting by Zhixin Liu, Xuanlin Liu, Sihan Xu, Yaqiong Qiao, Ying Zhang, Xiangrui Cai (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04247.txt.
Description
Zhixin Liu and five co-authors design a backdoor attack on time series forecasting that fires late and unevenly. A backdoor hides a trigger in training data so a model misbehaves on cue. Existing attacks force the trigger and its target pattern to line up in time and across variables, which real settings often cannot promise.
Their framework, TDBA, injects triggers that encode where the target pattern should appear later. Activation can then land at any forecast position and differ by variable. One module builds position-guided triggers from smoothed Gaussian priors, and another weights training by trigger completeness, pattern coverage, and time offset. On real data sets TDBA beats existing baselines on effect while staying stealthy, and removal studies confirm the design can be controlled and holds up well.
Catalog record
| CREATOR | Zhixin Liu, Xuanlin Liu, Sihan Xu, Yaqiong Qiao, Ying Zhang, Xiangrui Cai |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-04247