STA: A Split Temporal Adversarial Method for Evaluating the Robustness of Multimodal Reinforcement Learning Driving Policies
STA: A Split Temporal Adversary Method for Evaluating Robustness of Multimodal RL-based Driving Policies
The authors propose the STA attack; in CARLA Town02, the collision rate against the PPO policy is 35.6% (ADV is 16.8%), with 17.8% fewer attack steps, and it generalizes to a VLM policy (62%).
black-box threat model: the attacker can observe the victim policy's output actions but cannot see its action distribution or parameters; the attacker creates cross-modal temporal inconsistency by delaying the observation of a selected modality, with a limited budget (B=30), and the goal is to induce failures such as collisions.
- Multimodal reinforcement learning driving policies rely on cross-modal temporal alignment, and the authors consider temporal inconsistency a core vulnerability; existing attacks are insufficient in generality, realism, and adaptivity, making it difficult to systematically evaluate this vulnerability.
- Proposes Split Temporal Adversary (STA): uses a history buffer to replace the current observation of a selected modality with an n-step delayed observation to achieve unified temporal perturbation, then uses a separate decision mechanism to jointly optimize attack timing and modality selection under a budget constraint, trained with PPO.
- In the CARLA Town02 experiments, STA's collision rate against PPO is 35.6% (the strongest baseline ADV is 16.8%) with 17.8% fewer attacks; against DARRL, ADV is instead stronger (32.2% vs 18.6%); against the VLM policy, the collision rate reaches 62%.
- Future directions listed by the authors: more restricted black-box query settings, learnable delay policies, extension to more modalities such as LiDAR, and real-world and multi-agent environments. The experiments cover only CARLA Town02 dense traffic, fixed clear weather, and 4 victim policies.
- Evaluated models
- PPOSACDARRLVLM-based driving policy [1]
- Benchmarks
- CARLA 0.9.13 Town02 dense traffic
- Metrics
- CR (collision rate)ANA (average number of attacks)
The paper proposes a temporal adversarial attack method, STA (Split Temporal Adversary), against multimodal reinforcement learning driving policies, exploiting cross-modal temporal inconsistency as a vulnerability to induce decision failures. The method models the multimodal temporal attack as a unified decision process; under the realistic setting of policy awareness and limited budget, it uses delayed observation of historical states, avoids modality-specific noise, and jointly optimizes attack timing and modality selection, while using the victim policy's clean actions to identify sensitive decision patterns; the paper also provides a theoretical analysis of the induced distribution shift and principles for modality selection. In CARLA simulator experiments, STA causes a collision rate against PPO more than twice that of the strongest baseline, reduces the number of attack attempts by 17.8%, and generalizes to VLM-based driving policies. The authors state that these results show temporal consistency is a key vulnerability and that temporal robustness is critical to autonomous driving safety.
Paper analysis
What problem does this paper attempt to solve?
The authors consider cross-modal temporal inconsistency in multimodal RL driving policies to be the core vulnerability, and note that existing attacks are difficult to use for systematic evaluation; they propose STA.
How to systematically evaluate the robustness of multimodal reinforcement learning (MRL) driving policies under cross-modal temporal inconsistency.
- Problem setting: The authors argue that the effectiveness of MRL driving policies relies on the temporal alignment of heterogeneous sensor observations, and that any misalignment, delay jitter, or asynchrony pollutes the fused representation and harms decision reliability, so temporal inconsistency itself constitutes a core vulnerability.
- Shortcomings of existing methods (according to the paper): most attacks rely on modality-specific perturbations (e.g., image noise, LiDAR point perturbations) and lack unified formalization; they ignore victim policy feedback, making the threat model unrealistic; temporal perturbation methods use random modality selection or fixed injection schedules and lack adaptive optimization of attack timing and modality selection.
- Core observation: the synchronization mechanism only checks whether timestamp headers are within tolerance τ (R-RTM) and does not check semantic content, so an attacker can make the synchronizer pair observations from different time steps, causing semantic-relative temporal misalignment (S-RTM), and ∆sem can be arbitrarily large (Section III-A).
- Proposed in the paper: Split Temporal Adversary (STA), which formalizes multimodal temporal attacks as a unified decision process, jointly optimizes attack timing and modality selection under policy-aware and budget constraints, and provides a theoretical analysis of distribution shift. The authors claim STA is the first adversarial framework that unifies temporal perturbation across heterogeneous modalities under realistic attack constraints and adaptively optimizes attack timing and modality selection.
- Threat model:
- Goal: induce safety-critical failures (e.g., collisions).
- Knowledge: black-box—the attacker can observe the victim policy's output actions but cannot see its action distribution or parameters (Section VI-A).
- Capability: replace the current observation of a selected modality with an n-step delayed observation via a history buffer; the delay is bounded within horizon n; the attack budget is limited (B=30).
- Victim systems: fixed multimodal RL driving policies in CARLA (PPO, SAC, DARRL) and a VLM-based driving policy; policies are frozen during evaluation.
What related research exists?
Related work falls into three groups: DRL adversarial attacks, multimodal driving robustness, and temporal attacks; the authors claim STA fills the gap of adaptive policy-aware budget-constrained temporal attacks.
Adversarial attacks on deep reinforcement learning
- Huang et al. [8]—early work, in which small observation perturbations in Atari environments can degrade policy performance; subsequent work [9][10][24] uses policy-sensitive objectives to improve attack effectiveness, e.g., [10] raises the success rate by 18.7% through policy-sensitive targeting.
- [25]–[27]—model attacker-victim interaction as a sequential decision process. The paper points out that these DRL attacks perturb observations independently step by step, ignore long-term temporal dependencies, and assume access to all modalities; work considering temporally correlated perturbations [28][29] remains limited to a single modality.
Robustness of multimodal driving policies
- [4][5]—unimodal perturbations can propagate through fusion networks and degrade downstream policy performance.
- [11][13]–[15]—formalize adversarial driving as a constrained optimization or reinforcement learning problem to induce unsafe behavior; [30]–[32] propose robustness evaluation and certification methods. The paper points out that these works mainly focus on spatial or scenario-level perturbations and do not explicitly consider temporal inconsistency across heterogeneous modalities.
Temporal attacks and latency vulnerability
- Chronos [18]—the authors state it first shows that software-induced timing delays can degrade perception and planning; DejaVu [16]—a single-frame delay reduces camera-LiDAR detection mAP by 88.5%, but the paper points out that it does not evaluate the downstream impact on driving decisions.
- SlowPerception [37]—reaches a 97% collision rate, but the paper points out that its delay scheduling is predetermined rather than adaptively optimized; [36]—a study of optimal attack scheduling provides a theoretical basis for adaptive scheduling. The paper points out that existing temporal attacks are mostly open-loop, fixed, or heuristic delays, lacking policy awareness and budget constraints.
Baselines and benchmarks: The experiments compare Clean, RTA (random temporal attack [37]), RTAM (all-modality random temporal attack [16]), and ADV (PGD-based adaptive attacker [26]), evaluated on CARLA 0.9.13 Town02 with dense traffic.
Positioning: The authors claim that no existing work addresses the temporal attack optimization of jointly selecting attack timing and delayed modality under adaptive, policy-aware, budget constraints, and that STA directly fills this gap.
How does the paper address this problem?
STA contains two mechanisms: unified multimodal temporal perturbation (history-buffer delayed replacement) and separate multimodal adversarial decision-making (PPO optimizes triggering and modality selection).
Overall approach: multimodal temporal attacks are modeled as a single-agent POMDP, and Split Temporal Adversary (STA) learns, under a budget constraint, when to attack and which modality to attack, with the perturbation uniformly formulated as "replace the current observation with a historical observation."
Problem setup
- The attacker-victim interaction is modeled as a POMDP M = ⟨ S, O, A, T, R, γ ⟩, where the victim is part of the environment and the attacker learns a policy to maximize discounted return.
- Assumptions: the driving policy is fixed during evaluation; the temporal delay is bounded within horizon n; all sensors run at a fixed synchronized rate; the attack budget B is limited.
Unified multimodal temporal perturbation mechanism
- History buffer: a ring buffer Ht is maintained to store the most recent n complete multimodal observation tuples; all modalities share the same temporal store, with no modality-specific buffers needed.
- Delayed replacement: when an attack is triggered, for the selected modality m, the current observation is replaced with the historical observation from n steps earlier, i.e., õt(m) = ot−n(m); no additive noise or modality-specific transformation is introduced, and the attack effect comes solely from cross-modal temporal misalignment at the fusion layer.
- Fixed delay: a fixed n-step delay is used (taking the oldest entry in the buffer); the paper states this provides a stable, reproducible perturbation strength; the delayed observation is itself a legitimate sensor reading, preserving intra-modality semantic consistency.
Split multimodal adversarial decision mechanism
- Policy-aware observation: the adversarial observation is ôt = (ot, uclean_t, bt), where uclean_t is the victim policy's clean action (estimating its local decision tendency) and bt is the remaining budget.
- Split action space: the action at = (atrig_t, amod_t); the trigger branch decides whether to attack when atrig_t > 0 and bt > 0, and the modality branch selects the target modality from {1, 2}; the paper states this decomposition lets the two decisions be optimized independently.
- Budget and reward: the budget decreases by 1 per attack; the reward is sparse, with Rc awarded when a collision occurs (including within a T-step window after an attack), and 0 otherwise.
- Gated optimization objective: Ladv = Ltrig + 1(ãtrig_t = 1)Lmod; the modality selector is updated only when an attack is executed, addressing the credit assignment problem of the split action space.
Optimization and procedure
- The adversarial policy is optimized with PPO (800,000 training steps), with the victim policy frozen; a trajectory truncation strategy is used (following Fan et al. [21]): the episode is truncated when the budget is exhausted, and if the last attack causes a collision the episode terminates immediately and the trajectory is kept; otherwise the environment continues for at most T steps to capture delayed effects.
Theoretical analysis (Section V)
- Distribution shift lower bound (Proposition V.A): under a k-step delay, the total variation between the perturbed distribution and the nominal distribution is at least the expected divergence between the distributions of the attacked modality.
- Perturbation propagation bound (Proposition V.B): under an L-Lipschitz policy, the action deviation is at most L times the observation deviation, and under an additive metric only the attacked modality contributes.
- Modality selection principle (Corollary V.A): modalities with larger local sensitivity ‖Jm‖ and greater temporal variation are more influential; the authors consider numerical modalities often more sensitive because they directly affect the control signal.
What experiments did the paper conduct?
On CARLA Town02, STA's collision rate against PPO is 35.6% (ADV 16.8%) with 17.8% fewer attacks; against DARRL, ADV is instead stronger; on the VLM policy, 62%.
Experimental setup
- Victim policies: PPO (trained for 440,000 steps), SAC (235,000 steps), DARRL (the authors' reproduction in CARLA [22]), all trained on clean observations and frozen; plus the VLM-based driving policy of [1].
- Environment: CARLA 0.9.13, Town02, dense traffic, fixed clear weather; the victim receives RGB images, acceleration, and waypoint state, and outputs continuous steering and throttle.
- Threat model: black-box; the attacker can observe the victim's output actions but cannot see the action distribution or parameters.
- STA training: PPO trained for 800,000 steps; default delay horizon n=10, history buffer length 10, attack budget B=30, episode length at most 50 steps.
- Baselines: Clean, RTA (fixed n-step delay on a random modality at a random time step [37]), RTAM (random delay added to both modalities [16]), ADV (PGD perturbation with adaptive attack timing [26]).
- Metrics and repetitions: CR (collision rate, the percentage of episodes terminated by collision) and ANA (average number of attacks per episode); 5 independent seeds with 50 evaluation episodes each, reporting the mean and variance across seeds.
Main results
Table I (Town02 dense traffic, CR % / ANA, mean ± variance over 5 seeds):
Swipe horizontally to read the table
Method PPO CR PPO ANA SAC CR SAC ANA DARRL CR DARRL ANA Clean 6.8 0 2.8 0 2.8 0 RTA 4.0 29.524 4.0 29.820 4.4 29.760 RTAM 12.0 28.864 5.2 29.660 6.8 29.532 ADV 16.8 29.396 16.8 29.532 32.2 28.932 STA 35.6 24.168 18.2 25.628 18.6 26.124 - The authors state that STA is the strongest attack on PPO and SAC with fewer interventions: on PPO, CR rises from ADV's 16.8% to 35.6%, and ANA falls from 29.396 to 24.168 (17.8% fewer); on SAC the advantage is smaller (18.2% vs 16.8%).
- Exception: DARRL is more vulnerable to ADV (32.2%) than to STA (18.6%); the authors conclude from this that robustness to spatial perturbations does not imply robustness to temporal inconsistency, and evaluation should cover both threat models.
- The random temporal baselines nearly exhaust the budget but perform poorly, which the authors consider to support the value of policy-aware attack timing and modality selection.
Evaluation on the VLM-based driving policy
- Setup: same Town02 configuration (B=30, n=10); STA delays image frames; the text instruction is fixed at episode initialization and is not attacked.
- Results (Table II): under Clean, RTA, RTAM, and ADV, CR is 0–2% (Clean 2.00%, RTA 0.00%, RTAM 2.00%, ADV 2.00%), while STA reaches 62.00% CR and ANA 26.98.
- Authors' interpretation: the authors consider that policy-aware temporal perturbation can expose failure modes beyond those of conventional RL policies, while noting that evaluation across more VLM architectures is left for future work.
Other ablations and analyses
- The paper does not report separate ablation or parameter analysis experiments; the theoretical results in Section V (distribution shift lower bound, Lipschitz propagation bound, modality selection principle) are not accompanied by experimental validation.
- Negative result: RTA's CR on PPO (4.0%) is lower than Clean (6.8%), and on the VLM policy RTA is 0.00%, also lower than Clean's 2.00%; the paper does not explain this.
What are some points that can be further explored?
The authors list four future directions; the experiments cover only the single CARLA Town02 scenario and 4 victim policies, and no ablation is reported.
Limitations and future directions noted by the authors
- Explore more restricted black-box settings and limited query access (Section VII).
- Adaptive, learnable delay policies (currently a fixed n-step delay) (Section VII).
- Extend to richer sensor modalities such as LiDAR, radar, and GPS (Section VII).
- Study the applicability of STA in real-world and multi-agent driving environments (Section VII).
- Evaluation across more VLM architectures is left for future work (Section VI-C).
Scope of the experiments
- 4 victim policies tested: PPO, SAC, DARRL (reproduced by the authors), and one VLM-based driving policy (Section VI-A, VI-C).
- Environment is only CARLA 0.9.13 Town02 dense traffic, fixed clear weather; episodes are at most 50 steps (Section VI-A).
- Modalities are RGB images plus numerical state (speed, waypoints); in the VLM experiments only image frames are delayed, and text instructions are not attacked (Section VI-A, VI-C).
- Each condition is 5 seeds with 50 evaluation episodes each; the paper does not report ablation experiments, parameter sweeps of the delay horizon n or the budget B, or safety metrics other than attack success rate.
- The theoretical analysis (Section V) is not accompanied by experimental validation; the paper does not report adversarial evaluations against defense methods.
Summarize the main content of the paper
STA unifies multimodal temporal attacks as delayed replacement plus separate decision-making; on CARLA, its 35.6% collision rate against PPO exceeds the strongest baseline by more than twofold, and it generalizes to the VLM policy.
An attack paper targeting robustness evaluation of multimodal reinforcement learning driving policies.
- Problem: The authors argue that MRL driving policies rely on cross-modal temporal alignment, that the synchronization mechanism only checks timestamps and not semantics, and that temporal inconsistency constitutes a core vulnerability; existing attacks lack unified formalization, policy awareness, and adaptive optimization.
- Method: STA uses a history buffer to replace the current observation of a selected modality with a fixed n=10-step delayed observation (modality-agnostic, no additive noise), and uses a separate decision mechanism to jointly optimize attack triggering and modality selection under budget B=30, trained with PPO, with a gating objective handling credit assignment; theoretically it gives a distribution shift lower bound and a policy-Jacobian-based modality selection principle.
- Main results (CARLA Town02, Table I): STA's collision rate against PPO is 35.6%, the strongest baseline ADV is 16.8%, with 17.8% fewer attacks (ANA 24.168 vs 29.396); against SAC, 18.2% vs 16.8%.
- Exception: DARRL is more vulnerable to ADV (32.2%) than to STA (18.6%), from which the authors conclude that robustness to spatial perturbation does not imply temporal robustness.
- Generalization: On the VLM-based driving policy, STA reaches 62.00% collision rate, while the baselines are only 0–2% (Table II).
- Authors' conclusion: Disrupting cross-modal synchronization alignment is a key vulnerability, and temporal robustness is essential for safe autonomous driving; future directions include more restricted query settings, learnable delays, more modalities, and real-world evaluation.