acceptodds
Under review as a conference paper at ICLR 2027

Can a Backdoor Span Source and Wireless Domains? A Cross-Domain Backdoor Attack on Semantic Communication

13 views

Abstract

Artificial intelligence is reshaping wireless communications towards intelligent paradigms. Semantic communication (SemCom) exemplifies this shift by leveraging deep neural networks (DNNs) to transmit semantic representations, substantially improving communication efficiency. However, this reliance on DNNs exposes SemCom to backdoor attacks. Its DNN-based architecture facilitates backdoor implantation in the source domain, while the open nature of the wireless channel introduces another attack surface. Existing backdoor attacks predominantly focus on a single source domain. This raises a fundamental question: can a single backdoor span the source and wireless domains to enable a stealthier and more effective attack? Motivated by this, we propose XComBo, the first cross-domain backdoor attack against SemCom. It uses sub-trigger pairs with a learnable source-domain pattern and a wireless-domain frequency-structured signal, where only matched pairs activate the backdoor. Joint optimization on a surrogate model coordinates the sub-triggers for reliable activation without prior knowledge of the victim model, while energy constraints preserve benign behavior under incomplete trigger conditions. We further extend the framework to support multi-target attacks. Experiments show an average PSNR improvement of 12.47 dB over the strongest baseline under backdoor activation while preserving benign transmission. XComBo also generalizes across architectures and datasets and remains robust against several defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
semantic communication, wireless communication, backdoor attacks, cross-domain attacks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_aZFUOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a cross-domain backdoor attack method against semantic communication systems, named XComBo, which superimposes perturbations that require matched activation in both the source domain and the channel domain to achieve a covert backdoor attack on semantic communication systems. The authors tested the effectiveness of the proposed method on three common semantic communication system models across four datasets.

Critical strengths
  1. The authors claim that the proposed method is the first semantic communication attack method that distributes backdoor trigger conditions across two heterogeneous attack surfaces, the source and the wireless channel, and it requires no substitute joint optimization framework based on prior knowledge of the victim model.

  2. The FDM-inspired multi-sinusoid signal trigger with channel pre-compensation, the FPR energy constraint, and the surrogate-based joint optimization are technically sensible and carefully derived. The energy constraints are explicitly tied to the goal of preserving benign behavior under incomplete trigger conditions.

Critical weaknesses
  1. Section 3.1 cites one reference for timing synchronization, yet any symbol timing or carrier frequency offset would rotate the optimized phases of the signal sub-trigger (Eq. 7) and corrupt the injection (Eq. 11). The authors carefully analyzed imperfect CSI in Appendix D.1, not the arguably more fragile synchronization assumption.

  2. The baseline set is limited, and the comparisons are inherently approximate: Only three adapted baselines are considered, and none is a source-domain backdoor attack on SemCom, although such attacks exist and are cited in the related work (e.g., Sagduyu et al. 2023; Xu et al. 2024; Zhou et al. 2024; Yang et al. 2026).

Questions for authors

1.Training-time wireless injection: Please clarify the assumed access for wireless injection during training-time poisoning: under what conditions can the adversary observe and interfere with the victim's training transmissions in a real system, and does the attack still succeed if the injection channel differs between training and inference?

2.Real-world evaluation: All experiments use a simulated Rician channel with idealized injection (y + λφs). Do the authors expect the attack to hold under real over-the-air conditions and would they consider validating it with a hardware or over-the-air test?

7 comments

We thank the reviewer for their constructive feedback on practical wireless conditions and hardware validation. We appreciate their time and patience in considering our further clarification, additional simulation and real over-the-air experiments.

Q1: Please clarify the assumed access for wireless injection during training-time poisoning: under what conditions can the adversary observe and interfere with the victim's training transmissions in a real system? (1/2)

Training-time injection access. Training-time poisoning does not require observing or interfering with physical wireless transmissions. For backpropagation-based end-to-end training, physical wireless channels generally cannot directly propagate the required channel gradients. SemCom systems therefore use differentiable simulated channel modules [A5–A7]. The assumed poisoning access covers selected training samples, reconstruction targets, and the simulated channel module, allowing the attacker to add the signal sub-trigger througth the training pipeline. Channel conditions and timing are specified within the simulation, so physical channel estimation and real-time synchronization are not required during poisoning. Adversary can obtain these capabilities in practical scenarios such as outsourced training [A8].

Inference-time capabilities. Inference stage does not require access to or control over the deployed victim model.

Channel estimation. The adversary can exploit co-location with the legitimate receiver to obtain highly correlated estimates of the transmitter-to-receiver channel under spatially correlated channel conditions [A1]. For the adversary-to-receiver channel, observations of receiver-originated transmissions support estimation through channel reciprocity [A2].

Timing synchronization. Timing synchronization is a relatively strong assumption in practical deployment, although commonly assumed in wireless simulations [A9–A11]. So we discuss two possible implementation approaches, and discusses XComBo's effectiveness under out-of-sync conditions: (a) Preamble detection and reactive transmission. The adversary detects the beginning of a transmitted packet’s preamble and responds by transmitting the backdoor signal. RFReact reports microsecond-scale operation [A3], while an FPGA-based USRP N210 framework reports a response of approximately 80 ns after signal detection [A4]. (b) Repeated signal transmission. The adversary can transmit multiple copies of the backdoor signal to accommodate potentially large temporal misalignment [A2]. This approach may leave a random relative offset δ\delta between the legitimate signal and the backdoor signal. Neither approach requires modifying the receiver’s model or inference pipeline.

Q1: Please clarify the assumed access for wireless injection during training-time poisoning: under what conditions can the adversary observe and interfere with the victim's training transmissions in a real system? (2/2)

To evaluate the remaining timing uncertainty, we study random timing offsets during poisoning and evaluate the models temporal misalignment:

Table R1. Random timing offsets results on WITT/GTSRB at SNR = 18 dB and FPR = 15 dB. Training and testing independently sample an shift for each image from the listed distribution, measured under PSNR (dB):

Offset distributionStrategyNMAOBOBDMS
(-4)--(+4)Joint38.6437.3938.1841.0737.05
(-4)--(+4)Random38.1334.1737.0938.6934.57
(0)--(L-1)Joint38.2236.7037.8841.6336.59
(0)--(L-1)Random37.9229.4836.8928.4930.06

XComBo retains BD PSNRs of 41.07/41.63 dB under local-ranged or full-ranged random shifts, while NM/AO/BO/MS preserve source reconstruction at 36.59–38.64 dB. These results show that the random timing offsets can be accommodated through training, without requiring victim-model access at inference.

References

[A1] Physical Layer Key Generation in 5G Wireless Networks. IEEE Wireless Communications, 2019.

[A2] Exploring Practical Vulnerabilities of Machine Learning-based Wireless Systems. USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2023.

[A3] RFReact: A Real-time Capable and Channel-aware Jamming Platform. ACM Conference on Wireless Network Security (WiSec), demo, 2011.

[A4] A Real-Time and Protocol-Aware Reactive Jamming Framework Built on Software-Defined Radios. ACM Workshop on Software Radio Implementation Forum (SRIF), 2014.

[A5] Deep Joint Source-Channel Coding for Wireless Image Transmission. IEEE Transactions on Cognitive Communications and Networking, 2019.

[A6] Deep Learning Enabled Semantic Communication Systems. IEEE Transactions on Signal Processing, 2021.

[A7] Deep Learning-Based End-to-End Wireless Communication Systems With Conditional GANs as Unknown Channels. IEEE Transactions on Wireless Communications, 2020.

[A8] BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain. arXiv:1708.06733, 2017.

[A9] Adversarial Attacks Against Shared Knowledge Interpretation in Semantic Communications. IEEE Transactions on Cognitive Communications and Networking, 2025.

[A10] Channel-Aware Adversarial Attacks Against Deep Learning-Based Wireless Signal Classifiers. IEEE Transactions on Wireless Communications, 2022.

[A11] Adversarial Attacks on Deep-Learning Based Radio Signal Classification. IEEE Wireless Communications Letters, 2019.

\

Q1: Does the attack still succeed if the injection channel differs between training and inference?

We additionally evaluate changes in channel distribution and Rician fading parameters without reoptimizing the model.

Table R2. Reconstruction PSNR of XComBo under different training and inference channels on WITT/GTSRB at SNR=18 dB and FPR=15 dB, measured under PSNR(dB).

Training channelInference channelNMAOBOBDMS
Rician K=1K=1Rician K=0.1K=0.138.2537.0037.8944.6536.91
Rician K=1K=1Rician K=1K=138.8837.5238.3345.1837.27
Rician K=1K=1Rician K=10K=1040.9839.5940.4446.0839.28
RayleighRician K=0.1K=0.137.9136.5237.4144.2736.27
RayleighRician K=1K=138.4836.9537.8244.6436.66
RayleighRician K=10K=1040.4938.8839.8345.3838.43
RayleighRayleigh38.0236.6037.3944.4936.29
RayleighAWGN40.7639.1340.0445.3638.66
AWGNRician K=0.1K=0.129.5428.7329.4633.7529.27
AWGNRician K=1K=131.8130.2631.3735.9430.48
AWGNRician K=10K=1043.3741.7942.6844.1541.30
AWGNRayleigh30.0829.1129.4034.4829.30
AWGNAWGN43.8442.2043.0744.1941.77

The Rayleigh-trained model maintains BD PSNR of 44.27–45.38 dB across the tested Rician, Rayleigh, and AWGN channels. The Rician-trained model maintains 44.65–46.08 dB across the tested KK values. The AWGN-trained model has lower target fidelity under stronger fading, but still obtains BD PSNR of 33.75–35.94 dB in the reported low-KK Rician and Rayleigh settings. Its NM PSNR also decreases under these fading channels, suggesting that the lower BD fidelity partly reflects degradation of the underlying communication model. The results show that target reconstruction can be retained despite differences between training and inference channels.

Q2: Real-world evaluation: All experiments use a simulated Rician channel with idealized injection (y + λφs). Do the authors expect the attack to hold under real over-the-air conditions and would they consider validating it with a hardware or over-the-air test?

We thank the reviewer for this suggestion. We conducted over-the-air validation using three USRP B210, configured as the legitimate transmitter, receiver, and adversarial transmitter, respectively. Experiments were conducted at a carrier frequency of 2.2GHz, without timing synchronization between the legitimate and adversarial transmitters.

Within the available time and equipment constraints, we implemented an OTA attack without channel estimation, as is also adopted in SemCom systems [H1-H2], using WITT/GTSRB model trained with random timing offset. This implementation enables direct evaluation of the cross-domain threat over a real wireless transmittion.

As shown in Table R3, BD reconstructed target images at 38.09--41.37 dB in PSNR, while individual triggers and mismatched pairs retained source-image reconstruction, consistent with simulation. Some degradationm of reconstruction quality is expected compared with transmission using channel equalization. The higher BD PSNR relative to simulation may reflect less severe timing misalignment during OTA testing (compared with full random misalignment in simulation) although we started legitimate and adversarial transmitter individually and manually without synchronization.

Table R3. Simulation OTA results on WITT/GTSRB. Target A–C denote the three attack targets. OTA values are mean ±\pm standard deviation across repeated recordings, calculated from each recording’s mean per-image PSNR.

Pattern triggerSignal triggerConfigurationSimulation PSNR ↑\uparrow (dB)OTA PSNR ↑\uparrow (dB)
----NM33.6030.65 ± 2.94
Target A--AO31.1429.78 ± 0.77
Target B--AO31.3928.73 ± 1.84
Target C--AO31.4928.71 ± 2.23
--Target ABO30.7831.69 ± 3.86
--Target BBO31.4831.38 ± 0.45
--Target CBO32.2929.59 ± 3.02
Target ATarget ABD30.1538.09 ± 0.51
Target BTarget BBD30.5741.37 ± 1.58
Target CTarget CBD29.4141.05 ± 2.96
Target ATarget BMS30.36 (averaged under all mismatch configurations)28.32 ± 1.47
Target BTarget CMS30.36 (averaged under all mismatch configurations)28.88 ± 2.05

references:

[H1] Deep Joint Source-Channel Coding for Wireless Image Transmission. IEEE Transactions on Cognitive Communications and Networking, 2019.

[H2] Backdoor Attacks and Defenses on Semantic-Symbol Reconstruction in Semantic Communications. IEEE International Conference on Communications (ICC), 2024.

Reviewer_38vDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to design a new backdoor attack against semantic communication systems. The idea is that only the matched source-image perturbation and a wireless signal trigger will enable such a backdoor attack. The triggers are jointly optimized on a surrogate model and subsequently used to poison victim training. Experiments cover three semantic communication architectures and four image datasets, alongside ablations and evaluations against several defenses.

Critical strengths
  • The motivation and storyline are introduced clearly. It is a new attack setting for semantic communication with dual-key backdoors and surrogate-based trigger optimization, and channel per-compensation.

  • The generalization of the attack has been evaluated in detail. The authors evaluate the attack on the image reconstruction task across three SemCom systems and four datasets. It improves target-reconstruction PSNR by 12.47 dB.

Critical weaknesses
  • The contribution appears to lie primarily in extending established dual-key backdoor ideas to a source–wireless setting.

Although this extension to this new attack setting is potentially useful, the paper does not clearly establish which technical obstacles require the proposed joint optimization. Table 4 shows that random paired triggers already achieve substantial target-reconstruction fidelity, while joint optimization provides a smaller improvement over decoupled optimization. The paper would benefit from clearer evidence of the technical challenges and security insights specific to the wireless setting. The authors may also need to discuss the contribution compared to the existing dual-key multimodal backdoors.

  • The practical requirements of the threat model need clearer justification.

The attacker must modify source images and reconstruction targets during training, coordinate wireless trigger injection, estimate both relevant channels, and synchronize with the legitimate transmission. These capabilities are stated, but the paper does not sufficiently explain a realistic deployment scenario in which they are jointly available.

The wireless pre-compensation also uses the ratio of estimated channel gains, while the perturbation budget is imposed at the decoder input. This does not directly constrain attacker transmit power, which can become large when the attacker-to-receiver channel is weak.

  • The ablation study shows that the joint optimization may not be necessary.

Table 4 shows that random cross-domain triggers already achieve target PSNRs of 37.75–41.96 dB. Joint optimization improves over decoupled optimization by 0.57–2.34 dB. These results do not establish that joint optimization is essential for reliable cross-domain activation.

Questions for authors
  1. What realistic training or supply-chain scenario grants the attacker access to source images, reconstruction targets, and wireless injections? Does the reported 3% poisoning rate include all individual-trigger and mismatched-trigger training examples?

  2. Ablation study: How much of the matched-pair behavior comes from joint trigger optimization, and how much comes from training on incomplete and mismatched configurations? Could you provide an ablation separating these effects?

5 comments

We thank the reviewer for the thoughtful feedback. We appreciate their time and patience in reviewing the clarifications and additional evidence below, as these points requires a detailed response.

Q1/W2: What realistic training or supply-chain scenario grants the attacker access to source images, reconstruction targets, and wireless injections? Does the reported 3% poisoning rate include all individual-trigger and mismatched-trigger training examples? (1/2)

Thank you for raising this point. We hope the following clarification would help.

Training access and the role of the surrogate. Our threat model can be instantiated as outsourced training [V5]. The user specifies the SemCom architecture and training task, provides the training data, and receives a trained checkpoint for local deployment. The untrusted training party can modify selected source images and replace their reconstruction target, and add the signal sub-trigger through the simulated channel module (Note that physical wireless channels generally cannot directly propagate the required gradients.[V6–V8]). Consequently, training-time injection does not require observing or interfering with live over-the-air transmissions, channel estimation and synchronization are inference-time requirements. These permissions arise from the same training role, and the delivered checkpoint retains the original model architecture.

Inference-time capabilities. Activation requires modifying the source images and transmitting the wireless sub-trigger. It does not require control of the victim model or changes to its inference implementation.

Channel estimation. The adversary can exploit co-location with the legitimate receiver to obtain correlated transmitter-to-receiver channel estimates[V1]. For the adversary-to-receiver channel, receiver-originated transmissions support estimation through channel reciprocity [V2]. These communication-side observations do not require access to the victim model.

Timing Synchronization. Timing synchronization is a relatively strong assumption in practical deployment, although commonly assumed in wireless simulations [V9–V11]. So we discuss two possible implementation approaches, and discusses XComBo's effectiveness under out-of-sync conditions:

(a) Preamble detection and reactive transmission. The adversary can detect the beginning of a transmitted packet’s preamble and respond by transmitting the backdoor signal. Reactive-radio platforms report microsecond-scale responses [V3] and in an FPGA-based USRP N210 implementation, approximately 80 ns after signal detection [V4].

(b) Repeated signal transmission. The adversary can transmit multiple copies of the backdoor signal to accommodate potentially large temporal misalignment [V2]. This approach may leave a random timing offset δ\delta between the legitimate signal and the backdoor signal. Neither approach requires access to or control over the deployed victim model.

Q1/W2: What realistic training or supply-chain scenario grants the attacker access to source images, reconstruction targets, and wireless injections? Does the reported 3% poisoning rate include all individual-trigger and mismatched-trigger training examples? (2/2)

To evaluate the remaining timing uncertainty, we study random timing offsets during poisoning and evaluate the models temporal misalignment:

Table R1. Random timing offsets results on WITT/GTSRB at SNR = 18 dB and FPR = 15 dB. Training and testing independently sample an shift for each image from the listed distribution, measured under PSNR (dB)

Offset distributionStrategyNMAOBOBDMS
(-4)--(+4)Joint38.6437.3938.1841.0737.05
(-4)--(+4)Random38.1334.1737.0938.6934.57
(0)--(L-1)Joint38.2236.7037.8841.6336.59
(0)--(L-1)Random37.9229.4836.8928.4930.06

As shown in Table R1, XComBo retains BD PSNRs of 41.07/41.63 dB under local-ranged or full-ranged random shifts, while NM/AO/BO/MS preserve source reconstruction at 36.59–38.64 dB. These results show that XComBo remains effective when benign and adversarial transmitters are out-of-sync, without requiring victim-model access at inference.

The surrogate decouples trigger construction from victim-model poisoning. Before receiving a particular user's training task, the attacker can construct triggers using a surrogate and public data; after receiving the task, it uses the frozen triggers to poison the victim model without optimizing them anew for each architecture. The surrogate therefore supports advance optimization and reuse.

The 3% poisoning rate. The reported γ=0.03\gamma=0.03 is the fraction of training-source images selected for poisoning. We first select this subset, then derive its individual-trigger, matched-trigger, and mismatched-trigger configurations. Figure 5 additionally reports attack performance as this source-image fraction is reduced.

References

[V1] Physical Layer Key Generation in 5G Wireless Networks. IEEE Wireless Communications, 2019.

[V2] Exploring Practical Vulnerabilities of Machine Learning-based Wireless Systems. USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2023.

[V3] RFReact: A Real-time Capable and Channel-aware Jamming Platform. ACM Conference on Wireless Network Security (WiSec), demo, 2011.

[V4] A Real-Time and Protocol-Aware Reactive Jamming Framework Built on Software-Defined Radios. ACM Workshop on Software Radio Implementation Forum (SRIF), 2014.

[V5] BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain. arXiv:1708.06733, 2017.

[V6] Deep Joint Source-Channel Coding for Wireless Image Transmission. IEEE Transactions on Cognitive Communications and Networking, 2019.

[V7] Deep Learning Enabled Semantic Communication Systems. IEEE Transactions on Signal Processing, 2021.

[V8] Deep Learning-Based End-to-End Wireless Communication Systems With Conditional GANs as Unknown Channels. IEEE Transactions on Wireless Communications, 2020.

[V9] Adversarial Attacks Against Shared Knowledge Interpretation in Semantic Communications. IEEE Transactions on Cognitive Communications and Networking, 2025.

[V10] Channel-Aware Adversarial Attacks Against Deep Learning-Based Wireless Signal Classifiers. IEEE Transactions on Wireless Communications, 2022.

[V11] Adversarial Attacks on Deep-Learning Based Radio Signal Classification. IEEE Wireless Communications Letters, 2019.

Q2: Ablation study: How much of the matched-pair behavior comes from joint trigger optimization, and how much comes from training on incomplete and mismatched configurations? Could you provide an ablation separating these effects?

We sincerely thank the reviewer for this question. Our ablation shows that incomplete/mismatched supervision plays a key role in maintaining matched-pair selectivity, while joint optimization improves target fidelity during backdoor activation.

For each strategy, we compare poisoning under all five configurations (NM/AO/BO/BD/MS) with normal behavior and backdoor activation supervision (NM/BD). Results are shown in Table R2:

Table R2. Supervision ablation on GTSRB at SNR = 18 dB and FPR = 15 dB. PSNR values (dB) are mean ± standard deviation across architectures, except AO, BO, and MS under NM/BD supervision, which list ADJSCC/WITT/MambaJSCC separately. BD references the designated target; all other configurations reference the source image.

SupervisionStrategyNMAOBOBDMS
NM/AO/BO/BD/MSJoint37.84±1.6036.89±1.2737.44±1.5044.63±0.3836.62±1.13
NM/AO/BO/BD/MSDecoupled37.84±1.6536.89±1.2837.46±1.5942.88±0.9036.66±1.22
NM/AO/BO/BD/MSRandom37.59±1.5236.12±0.9536.64±1.3039.87±2.3835.60±0.74
NM/BDJoint38.54±1.7935.73 / 9.17 / 9.159.19 / 37.51 / 37.7247.36±0.639.17 / 9.16 / 9.13
NM/BDDecoupled38.53±1.8935.70 / 9.16 / 9.159.18 / 37.15 / 37.6447.35±1.249.15 / 9.15 / 9.14
NM/BDRandom38.42±1.7035.72 / 9.82 / 9.299.17 / 36.91 / 36.8646.78±1.299.14 / 9.22 / 9.22

Removing AO/BO/MS supervision causes all three strategies to develop single-domain shortcuts despite high matched-target fidelity. BO reconstructs to correspoinding backdoor targets at 47.09–48.40 dB on ADJSCC, while AO reaches 31.49–48.29/39.61–46.51 dB on WITT/MambaJSCC. MS outputs likewise follow one sub-trigger’s target, showing that the optimized triggers do not by themselves preserve paired activation in this ablation. (note that results for AO/BO/MS in Table R2 measures PSNR referencing the source images, not the backdoor target used above) Under complete supervision, all three strategies maintain paired activation, while Joint further improves BD PSNR by an average of 1.75/4.76 dB over Decoupled/Random.

Reviewer_2nKhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes XComBo, a cross-domain backdoor attack against semantic communication systems. The method combines a learnable source-domain pattern with a frequency-structured wireless-domain signal, requiring a matched pair of sub-triggers to activate the backdoor while preserving benign behavior under incomplete or mismatched trigger conditions. The authors develop a surrogate-based joint optimization framework and further extend the approach to multi-target attacks. Experiments across multiple architectures and datasets demonstrate attack effectiveness, cross-architecture generalization, and robustness under various communication conditions.

Critical strengths

The paper addresses a security concern in semantic communication systems, with a clear motivation. The proposed framework is coherent and well designed, integrating composite triggering, joint optimization, and energy constraints to support stealthy and multi-target attacks. The experimental evaluation is comprehensive, covering multiple architectures, datasets, and communication conditions, with convincing results supported by ablation studies and additional analyses.

Critical weaknesses
  1. Table 4 shows that random or decoupled sub-triggers already achieve relatively strong attack performance, while the additional gains from joint optimization are modest on some architectures. Further clarification of its specific contribution relative to model poisoning would be helpful.

  2. The proposed attack requires coordinated cross-domain injection, channel estimation, and timing synchronization, which may introduce practical challenges. In particular, how is the received semantic feature power required by FPR determined without prior knowledge of the victim model?

  3. Although the defense experiments are encouraging, SCALE-UP still achieves an AUC of 0.80, and some mitigation methods substantially reduce attack effectiveness at the cost of benign performance. A more nuanced discussion of these results would help clarify the robustness and limitations of the proposed attack.

Questions for authors
  1. Could the authors further disentangle the gains from joint trigger optimization and those from the poisoned trigger-response mapping, particularly in the settings where random or decoupled triggers already perform well in Table 4?

  2. What attacker knowledge and capabilities are required to estimate the received semantic-feature power used by FPR, and how feasible are channel estimation and timing synchronization when the victim model is not accessible?

  3. How should the SCALE-UP AUC of 0.80 and the attack-success versus benign-performance trade-offs under mitigation be interpreted when assessing the practical detectability and robustness of XComBo?

6 comments

We thank the reviewer for the careful reading and constructive questions. We appreciate his/her time and patience in considering our clarifications and additional experiments on optimization, attack feasibility, and defenses.

W1/Q1: Could the authors further disentangle the gains from joint trigger optimization and those from the poisoned trigger-response mapping, particularly in the settings where random or decoupled triggers already perform well in Table 4?

We thank the reviewer for this thoughtful question. Poisoning enables the victim to learn the target mapping, while joint sub-trigger optimization improves target fidelity under activation requirements and random timing offsets.

Ablation on supervision methods. We select two supervision settings: all configurations (NM/AO/BO/BD/MS) and configurations with normal behavior and backdoor activation (NM/BD). Results are shown in Table R1.

Table R1. Supervision ablation on GTSRB at SNR = 18 dB and FPR = 15 dB. PSNR values (dB) are mean ± standard deviation across architectures, except AO, BO, and MS under NM/BD supervision, which list ADJSCC/WITT/MambaJSCC separately. BD references the designated target, and all other configurations reference the source image.

SupervisionStrategyNMAOBOBDMS
NM/AO/BO/BD/MSJoint37.84±1.6036.89±1.2737.44±1.5044.63±0.3836.62±1.13
NM/AO/BO/BD/MSDecoupled37.84±1.6536.89±1.2837.46±1.5942.88±0.9036.66±1.22
NM/AO/BO/BD/MSRandom37.59±1.5236.12±0.9536.64±1.3039.87±2.3835.60±0.74
NM/BDJoint38.54±1.7935.73 / 9.17 / 9.159.19 / 37.51 / 37.7247.36±0.639.17 / 9.16 / 9.13
NM/BDDecoupled38.53±1.8935.70 / 9.16 / 9.159.18 / 37.15 / 37.6447.35±1.249.15 / 9.15 / 9.14
NM/BDRandom38.42±1.7035.72 / 9.82 / 9.299.17 / 36.91 / 36.8646.78±1.299.14 / 9.22 / 9.22

The results show that poisoning learns the target mapping, while joint optimization improves target fidelity during activation. Joint achieves the highest mean BD PSNR on all three architectures, averaging 1.75/4.76 dB above Decoupled/Random. Removing AO/BO/MS supervision preserves high BD fidelity across all three strategies but introduces single-domain shortcuts. BO reconstruct to the backdoor target at 47.09–48.40 dB on ADJSCC, while AO achieve 31.49–48.29 dB on WITT and 39.61–46.51 dB on MambaJSCC (note that results for AO/BO/MS in Table R1 measures PSNR referencing the source images, not the backdoor target used above).

Random timing offsets. Under random timing offsets, Joint also outperforms Random in target fidelity and source reconstruction, as shown in Table R2.

Table R2. Random timing offsets results on WITT/GTSRB at SNR = 18 dB and FPR = 15 dB. Training and testing independently sample an shift for each image from the listed distribution, measured under PSNR (dB):

Offset distributionStrategyNMAOBOBDMS
(-4)--(+4)Joint38.6437.3938.1841.0737.05
(-4)--(+4)Random38.1334.1737.0938.6934.57
(0)--(L-1)Joint38.2236.7037.8841.6336.59
(0)--(L-1)Random37.9229.4836.8928.4930.06

Joint’s BD advantage over Random grows from 2.38 dB under local-ranged shifts to 13.14 dB under full-ranged shifts. Under full-ranged shifts, Joint also improves AO/MS source-reconstruction PSNR by 7.22/6.53 dB, demonstrating benefits for both target activation and source preservation under random timing offsets.

W2/Q2: What attacker knowledge and capabilities are required to estimate the received semantic-feature power used by FPR?

For backpropagation-based end-to-end training, physical wireless channels generally cannot directly propagate the required gradients. DeepJSCC, DeepSC and other representative SemCom systems therefore train with differentiable simulated channel modules [P1, P5, P6]. Therefore the attacker can add the signal sub-trigger through this module without live over-the-air injection during poisoning. Channel conditions and timing are naturally available within the simulation, as is the intermediate feature power needed to set FPR.

At inference, access to the training pipeline is no longer available. Channel estimation and synchronization instead rely on adversary's observations, as discussed below, while random time shifts during poisoning improve tolerance to timing misalignment. For FPR, SemCom systems normalize encoder outputs before transmission [P1–P4], providing a known reference power of P0=1P_0=1. We therefore examine whether this fixed reference can replace the instantaneous power at the decoder input.

Table R3. Measured mean signal power at different positions in the SemCom pipeline and its deviation from the unit-power reference.

PositionMean powerDeviation from unit power
After transmitter power normalization1.0000.00 dB
Receiver input, before equalization1.018+0.08 dB
Receiver input, after equalization0.951−0.22 dB

Results shown that the measured mean powers remain close to the normalization reference. To assess the effect on attack performance directly, we compare calculation of the power of signal sub-trigger using the measured receiver-feature power (A) with the fixed reference P0=1P_0=1 (B).

Table R4. Reconstruction PSNR under measured-power and fixed-reference power calculation of signal sub-trigger on WITT/GTSRB at SNR =18=18 dB and nominal FPR =15=15 dB, measured in PSNR(dB).

Batch sizeCalculation referenceNMAOBOBDMS
64A: measured receiver-feature power38.8837.5238.3345.1837.27
64B: fixed reference P0=1P_0=138.8837.5238.3045.3337.25
1A: measured receiver-feature power39.7638.9139.2546.1938.63
1B: fixed reference P0=1P_0=139.7638.9139.2346.5438.62

Fixed-reference preserves target-reconstruction quality, while NM/AO are unchanged and BO/MS differ by at most 0.03 dB. These results are consistent with the energy-mismatch tolerance reported in Figure 4: accurate feature-power measurements can be used during training, whereas inference-time power calculation can use the known normalization reference without reducing attack fidelity in the evaluated setting.

References

[P1] Deep Joint Source-Channel Coding for Wireless Image Transmission. IEEE Transactions on Cognitive Communications and Networking, 2019.

[P2] Wireless Image Transmission Using Deep Source Channel Coding With Attention Modules. IEEE Transactions on Circuits and Systems for Video Technology, 2022.

[P3] WITT: A Wireless Image Transmission Transformer for Semantic Communications. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.

[P4] MambaJSCC: Adaptive Deep Joint Source-Channel Coding With Generalized State Space Model. IEEE Transactions on Wireless Communications, 2026.

[P5] Deep Learning Enabled Semantic Communication Systems. IEEE Transactions on Signal Processing, 2021.

[P6] Deep Learning-Based End-to-End Wireless Communication Systems With Conditional GANs as Unknown Channels. IEEE Transactions on Wireless Communications, 2020.

W2/Q2: How feasible are channel estimation and timing synchronization when the victim model is not accessible?

Channel estimation and synchronization without victim-model access. SemCom systems commonly include bidirectional data, control, or feedback links [L1]. Exploiting channel reciprocity, observations of receiver-originated transmissions can support estimation of the adversary-to-receiver channel [L2]. For transmitter-to-receiver channel estimation, the adversary can exploit co-location with the legitimate receiver. Under spatially correlated channel conditions, the adversary and receiver obtain highly correlated channel estimates from the legitimate transmitter's signals [L3]. These methods do not require access to the victim model.

Timing synchronization is a relatively strong assumption in practical deployment, although commonly assumed in wireless simulations [L4, L7, L8]. So we discuss two possible implementation approaches, and discusses XComBo's effectiveness under out-of-sync conditions:

(a) Preamble detection and reactive transmission. The adversary can detect the beginning of a transmitted packet's preamble and respond by transmitting the backdoor signal. RFReact reduces the delay between signal detection and reactive transmission from milliseconds to microseconds [L5], while an FPGA-based USRP N210 framework reports a response of approximately 80 ns after signal detection [L6].

(b) Repeated signal transmission. The adversary can also transmit multiple copies of the backdoor signal to accommodate potentially large temporal misalignment [L2]. This approach may leave a random timing offset δ\delta between the legitimate signal and the backdoor signal. Neither approach above requires access to or control over the victim model.

To evaluate the timing uncertainty under out-of-sync conditions, we study random timing offsets during poisoning and evaluate the models under temporal misalignment. Results are given in Table R2 in W1/Q1. XComBo retains BD PSNRs of 41.07/41.63 dB under local-ranged or full-ranged random shifts, while NM/AO/BO/MS preserve source reconstruction at 36.59–38.64 dB. These results show that XComBo remains effective when benign and adversarial transmitters are out-of-sync, without requiring victim-model access at inference.

References

[L1] A Survey on Semantic Communication Networks: Architecture, Security, and Privacy. IEEE Communications Surveys & Tutorials, 2025.

[L2] Exploring Practical Vulnerabilities of Machine Learning-based Wireless Systems. USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2023.

[L3] Physical Layer Key Generation in 5G Wireless Networks. IEEE Wireless Communications, 2019.

[L4] Adversarial Attacks Against Shared Knowledge Interpretation in Semantic Communications. IEEE Transactions on Cognitive Communications and Networking, 2025.

[L5] RFReact: A Real-time Capable and Channel-aware Jamming Platform. ACM Conference on Wireless Network Security (WiSec), demo, 2011.

[L6] A Real-Time and Protocol-Aware Reactive Jamming Framework Built on Software-Defined Radios. ACM Workshop on Software Radio Implementation Forum (SRIF), 2014.

[L7] Channel-Aware Adversarial Attacks Against Deep Learning-Based Wireless Signal Classifiers. IEEE Transactions on Wireless Communications, 2022.

[L8] Adversarial Attacks on Deep-Learning Based Radio Signal Classification. IEEE Wireless Communications Letters, 2019.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Discussion Summary and Updates

Dear Reviewers, ACs, SACs, and PCs,

We appreciate the continued attention to our submission and the constructive discussion. As the rebuttal will conclude soon, we would like to summarize the feedback received and the clarifications and additional experiments provided during rebuttal.

  • Reviewer 2nKh confirmed that our responses addresed their main concerns and clarified the contribution of joint optimization, the practical scenarios and assumptions of the attack, and the interpretation of the defense results.

  • Reviewer 38vD raised questions about the communication-specific contribution, the respective roles of joint optimization and configuration supervision, and the necessity of joint optimization. We provided clarifications and additional experiments addressing these questions.

  • Reviewer aZFU recognized that the added OTA experiments and analysis of carrier-frequency, phase, and timing offsets strengthened the validation and addressed concerns about real-world deployment and synchronization. Following their remaining concern about baseline coverage, we provided an additional quantitative comparison with a source-domain SemCom baseline.

Reviewer 38vD has not yet provided follow-up feedback, and we would welcome their assessment of the clarifications and additional evidence.

In response to the reviewers' feedback, we have:

  • Clarified the threat model and practical assumptions. (2nKh, 38vD, aZFU): We clarified outsourced training as a concrete scenario supporting the adversary's poisoning access, distinguished the adversary's training-time capabilities from inference-time access requirements, and explained the surrogate’s role in optimizing reusable triggers for different victim architectures before poisoning. We also further clarified how signal trigger strength can be determined at inference without access to the victim model. We also clarified the communication-specific challenges and the relationship to existing dual-key multimodal backdoors.

  • Clarified roles of joint sub-trigger optimization versus configuration supervision. (2nKh, 38vD): We carried out an ablation separating their contributions. Results clarify the role of supervision in preserving composite activation and the additional reconstruction fidelity gains from joint optimization during backdoor activation.

  • Added experiments on synchronization, channel changes, and real-world OTA attack. (38vD, aZFU): We expanded the discussion of synchronization approaches and added separate evaluations of carrier-frequency, phase, and timing offsets. We also evaluated random timing offsets during both poisoning and inference, and changes between training and inference channels. Beyond simulation, we conducted over-the-air (OTA) experiments using USRP B210, providing hardware evidence of the real-world cross-domain threat studied in our work.

  • Added a source-domain SemCom Baseline. (aZFU):Following the reviewer’s follow-up feedback, we added a quantitative comparison with a baseline derived from current source-domain SemCom backdoor attack. Source-domain methods require further adaptation because their original formulations do not directly address multi-target composite activation. Results confirm that our proposed method achieves outstanding performance.

  • Qualified the detection and mitigation conclusions. (2nKh): We added SCALE-UP evaluations at low false-positive operating points and discussed attack suppression together with its cost to normal reconstruction quality. These results provide a more qualified interpretation of stealthiness of our proposed attack.

We acknowledge that the OTA evaluation remains limited in synchronization characterization and implementation of channel equalization (aZFU). More extensive real-world evaluation would require additional equipment and experimental time beyond what is available during the rebuttal period. We regret that we cannot complete all these real-world experiments during rebuttal. Nevertheless, we suggest that the completed OTA experiments already provide direct evidence that the proposed cross-domain backdoor threat can manifest under real-world wireless transmissions.

Once again, thank you for your continued consideration and for helping us strengthen the paper.

Best regards, The Authors of Paper #14985

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.