BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting
summary
The gist
Multivariate long-term time series forecasting (MLTSF) models are increasingly deployed in critical domains such as climate, finance, and transportation.
In short
BadTime is a novel backdoor attack designed for Multivariate Long-Term Time Series Forecasting (MLTSF). It manipulates hundreds of future predictions toward a target pattern by injecting a subtle trigger. BadTime solves the challenges of trigger dilution over long horizons and sparse signals by using inter-variable correlations and mixed poisoned samples, achieving significant performance gains.
Key concepts
- Dilution of Local Triggers
- In long-term forecasting, the influence of a small, local trigger signal fades quickly as the prediction horizon increases. BadTime counters this by designing a distributed trigger that leverages inter-variable correlations and temporal lags, ensuring its effect persists across very long forecasts.
- Sparsity of Backdoor Signals
- Backdoor signals are often extremely weak or sparse when stealth constraints are applied. BadTime overcomes this by mixing label-similar and label-dissimilar poisoned samples. It also uses a decoupled training objective to explicitly boost the sparse backdoor signal, allowing for effective poisoning even at low rates.
- Lag-Aware Trigger Design
- The attack employs a distributed trigger strategy that is aware of temporal lags between different variables. This design ensures that the injected trigger's influence is maintained and relevant across long-range forecasts, rather than being lost due to the time difference between input features.
Terminology used across episodes
This episode discusses
- BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting · Paper Radio
- iTransformer: Inverted Transformers Are Effective for Time Series Forecasting
- Cyber Resilience Assessment of Unbalanced Distribution System Restoration under Sparse Load Forecasting Attacks
- Targeted Attacks on Timeseries Forecasting
- Mamba: Linear-Time Sequence Modeling with Selective State Spaces
- TSMixer: An All-MLP Architecture for Time Series Forecasting
The paper
BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting · Read on arXiv
Kunlan Xiang, Haomiao Yang, Meng Hao, Wenbo Jiang, Haoxin Wang, Shiyue Huang, Shaofeng Li, Yijing Liu, Ji Guo
University of Electronic Science and Technology of China · Singapore Management University · Sichuan University · Southeast University · Nanyang Technological University
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: I'm Nadia.
Elias: And I'm Elias, here with Priya.
Nadia: We're looking at the paper: "BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting".
Elias: Here's the gist of the paper. Multivariate long-term time series forecasting (MLTSF) models are increasingly deployed in critical domains such as climate, finance, and transportation.
Nadia: BadTime proposes the first effective backdoor attack tailored for MLTSF by manipulating hundreds of future predictions toward a target pattern by injecting a subtle trigger <ref:2508.04189#pg2> The Core Problem Addressed BadTime addresses two key challenges unique to MLTSF: (i) the rapid dilution of local triggers over long horizons, and (ii) the extreme sparsity of backdoor signals under stealth constraints <ref:2508.04189#pg4>.
Elias: Existing attacks are limited to MSTSF settings with prediction horizons of at most 12 steps, leaving MLTSF vulnerable <ref:2508.04189#pg8>.
Nadia: The paper initiates the study of backdoor attacks in MLTSF and reveals severe, previously unexplored backdoor vulnerabilities in widely adopted MLTSF models <ref:2508.04189#pg7>.
Paper summary: Nadia: So we're looking at this paper now "BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting." Essentially, the authors are saying that security for these long-term forecasting models is completely unstudied because they are being used in really critical areas like climate and finance.
Elias: Exactly. The main thesis is that they’ve created the first effective backdoor attack specifically designed for multivariate long-term time series forecasting, or MLTSF. They claim BadTime can manipulate hundreds of future predictions toward a target pattern just by injecting a subtle trigger into the historical data.
Priya: That sounds like it could be pretty serious if these models are used in real systems, because the paper mentions two major hurdles that plague MLTSF security: first, how local triggers get diluted over long horizons.
Nadia: Right. And second, there’s this extreme sparsity of backdoor signals when you try to keep the attack stealthy. They call that challenge (ii).
Elias: To tackle the dilution problem, BadTime designs a distributed, lag-aware trigger that uses things like inter-variable correlations and temporal lags to make sure the influence lasts over long-range forecasts.
Priya: That sounds complicated. So how do they actually build this trigger? What are these specific mechanisms they use?
Elias: They use three meticulous trigger designs. First, they employ Graph Attention Networks, or GATs, to model how variables influence each other and select which ones to inject the trigger fragments into.
Nadia: That’s where the inter-variable correlation comes in, right? So it’s not just one variable getting poisoned; it’s about how those relationships propagate.
Elias: Right. Second, they propose a lag-aware positioning strategy that uses temporal dependencies to place the trigger at timestamps aligned with peak delayed correlation, which ensures the activation happens where it matters most for long-term prediction.
Priya: So what about the sparsity issue? If you only poison a small part of the training data, how do you actually make sure that backdoor signal doesn't get completely drowned out by all the clean data?
Nadia: That’s where they use a hybrid selection strategy for poisoned samples. They mix two types of samples: label-similar samples whose ground-truth labels are close to the target pattern, and label-dissimilar samples whose labels are far away from it.
Elias: That mixing helps spread the signal out, which is smart for overcoming that sparsity challenge. And they use a decoupled backdoor training objective to explicitly up-weight that sparse signal.
Priya: So what does this decoupling actually look like in practice? What are those custom loss functions they’re using?
Elias: The model optimization step calculates the total loss as Lmodel, which involves three parts: Lc for clean performance, LVt p for learning the trigger-target mapping on poisoned samples, and LVbar t p to constrain predictions on non-target variables of poisoned samples to match their clean ground-truths.
Nadia: It sounds like they’re trying to balance getting a good forecast with making sure the backdoor doesn't mess up everything else. How does this actually compare to existing methods?
Elias: Extensive experiments show that BadTime significantly outperforms the state-of-the-art backdoor attack on time series forecasting. They extended the attackable horizon from at most twelve steps to seven hundred twenty steps, which is a sixty times improvement over what was previously possible <ref:2508.04189#pg3>.
Priya: Sixty times is a big jump. And what does that actually mean for someone who just looks at the error numbers?
Elias: It reduces the Mean Absolute Error on target variables by over fifty percent on average, and it boosts stealthiness by more than three times when anomaly detection is running. They also confirm that even when only one percent of the training samples are poisoned, the attack remains effective.
Nadia: That robustness against low poisoning rates is important for real-world deployment because you can't always guarantee perfect data purity in a critical system.
Elias: And it’s robust to variations in trigger length ratios, meaning the attack works even if the trigger isn't perfectly sized according to their initial design. It’s also effective across different pattern shapes.
Priya: I see they tested up to seventy percent of all variables and it stayed effective and stealthy then. That suggests a lot of versatility in how this attack can be set up against a complex MLTSF model.
Nadia: So, to wrap up this paper "BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting," the authors are presenting a method that tackles dilution and sparsity by designing a distributed, lag-aware trigger and using mixed samples to weight the backdoor signal.
Elias: It’s important to remember that they established this as the first effective attack tailored for MLTSF, which is a significant gap in security research.
Nadia: The implication here is that if these models are used for long-term forecasting in things like energy management or traffic control, we need to start thinking seriously about how to defend against this kind of poisoning.
Elias: BadTime moves the baseline for what’s possible with backdoor attacks on time series forecasting by drastically increasing the attack horizon and improving accuracy metrics.
Conclusion: Nadia: So we’re wrapping up our look at BadTime, which is this paper about finding a backdoor attack for multivariate long-term time series forecasting models like these really big ones used in finance or climate stuff.
Elias: Yeah, the authors are proposing this new technique called BadTime to manipulate hundreds of future predictions toward a specific pattern using just a subtle trigger. It’s the first one they’ve shown works well for this kind of complex time series data.
Priya: So what does that actually mean for us when we think about how much we can trust these long-term forecasts?
Nadia: Well, the core problem they tackled was two things in MLTSF. First, local triggers get diluted really fast over long horizons. And second, the backdoor signals are extremely sparse when you try to keep it stealthy.
Elias: BadTime handles that dilution by using inter-variable correlations and temporal lags to design a trigger that stays influential over long forecasts. That’s a distributed trigger, not just one spot in time or one variable.
Priya: And for the sparsity issue, they mix label-similar and label-dissimilar poisoned samples together, which helps make sure the backdoor signal actually gets learned even when you only poison a small percentage of the training data.
Nadia: The results they show are pretty striking. They extended the attackable horizon from maybe twelve steps up to seven hundred twenty steps, which is a sixty times improvement over what was possible before.
Elias: And on accuracy, they reduced the mean absolute error on the target variables by over fifty percent on average while keeping clean forecasting errors less than one percent. That’s a big jump in effectiveness.
Priya: So for someone just listening to this, what does this paper change about how we look at these models?
Nadia: It shows that the security of these long-term models against backdoor attacks was completely unexplored before, and BadTime is presented as the first effective one tailored specifically for MLTSF.
Elias: The authors are demonstrating a new way to design triggers that aren't just random noise, but are built to exploit the very structure of multivariate time series data.
Priya: It really highlights how sensitive these models are to subtle poisoning and how much they rely on those complex variable relationships for their long-term predictions.
Nadia: So we’ve seen how they build this attack and what the numbers show, but there’s still a lot to unpack about defending against it. Next up, we’re going to look at some of the defense techniques researchers are coming up with right now...
More episodes
- 2608.27496-ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection
- Daily Summary for 2026-11-19
- 2607.07085-Is Randomness Necessary for Adaptive Data Analysis?
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2609.10601-Threshold Choice, Not Sample Size, Bounds Trustless Verification of Nondeterministic Compound AI Workflows
- 2610.38786-Adversarial Process Disturbances at a Deployable Alarm Budget: Detector Evaluation on a Simulator and a Real Plant
- 2511.08660-Binary and Multiclass Cyberattack Classification on GeNIS Dataset
- 2604.27302-Static Attribution of Android Residential Proxy Malware Using Graph Kernels
- Daily Summary for 2026-11-18