iBANDA: A Blockchain-Assisted Defense System for Authentication in Drone-Based Logistics^†

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper introduces iBANDA—an AI- and blockchain-assisted defense system that provides simultaneous, low-latency authentication of drones and their attached payloads. The paper sounds novel and overall well-written. However, there are areas where improvements can be made.

1. In the abstract, how was the average end-to-end latency of 0.13 seconds obtained? What transmission protocol was adopted?
2. No quantitative analysis of attack resilience (e.g., Sybil/GPS spoofing).
3. Figure 4 overlooks MITM risks in the authentication flow.
4. It is highly suggested to improve the quality of some figures. For example, Fig. 4, 5, and 6. It seems like they were pasted from another book or article. These figures can be easily improved for changing their quality.
5. Some practical applications on the drone systems should be investigated and discussed in this paper. a) "Performance Analysis and Optimization Design of AAV-Assisted Vehicle Platooning in NOMA-Enhanced Internet of Vehicles," in IEEE Transactions on Intelligent Transportation Systems; b) "A AAV Swarm Authentication and Key Agreement Scheme Based on Latin Square Design," in IEEE Wireless Communications Letters (the author is free to cite recommendations).
6. Before the final document version, the authors should do a grammatical review of the paper. For example, the title of Figure 8 “based on the VisioDect [?]”.

Author Response

Comment 1: In the abstract, how was the average end-to-end latency of 0.13 seconds obtained?

Response: The authors recast the entire abstract section by adopting a structured format to capture all the necessary details of the manuscript as shown.

 

Comment 2: No quantitative analysis of attack resilience (e.g., Sybil/GPS spoofing).

 

Response: We thank the reviewer for this valuable observation. Although our current evaluation focused on authentication latency, throughput, and scalability, we agree that quantitative resilience testing against specific attacks such as Sybil and GPS spoofing would strengthen the work. The iBANDA framework is inherently designed to resist Sybil and spoofing threats through its Avalanche-based Proof-of-Stake consensus, repeated sub-sampling validation, and dual-layer (AI + blockchain) verification. In the revised manuscript, we included a resilience evaluation where the system is subjected to controlled Sybil and GPS spoofing scenarios in a simulated network environment. For Sybil attack analysis, we introduced a varying percentage of malicious validator nodes (e.g., 10–50%) and measure the system’s false acceptance rate, consensus time, and transaction rejection ratio. For GPS spoofing, we simulated false geolocation signals and quantify the detection success rate and authentication disruption time. These metrics will be compared against baseline DDS and other blockchain consensus models. This addition provided measurable evidence of iBANDA’s robustness under adversarial conditions and addresses the reviewer’s request.

Section 4.2.3 provides these details as shown.

 

Comment 4: Figure 4 overlooks MITM risks in the authentication flow.

 

Response: We appreciate the reviewer’s observation. While Figure 4 currently illustrates the high-level iBANDA authentication flow between the UAV, Ground Control Station (GCS), and the blockchain network, it does not explicitly depict security mechanisms addressing Man-in-the-Middle (MITM) attacks. Although the current design did not explicitly cover MITM attacks, with the End-to-End Mutual Authentication (Between UAV and GCS) and Dual-Layer Verification, the AI layer verifies behavioral patterns (e.g., UAV flight profile anomalies) that could detect an MITM-induced control hijack, while the blockchain layer verifies cryptographic integrity, thereby annulling its presence. As a future direction, the improved implementation of iBANDA framework will integrate multiple safeguards against MITM attacks and other attacks. All UAV–GCS–blockchain communications are secured via mutual TLS/DTLS 1.3 with blockchain-validated public keys, ensuring that both endpoints are authenticated before any data exchange. Session keys will be derived from blockchain-signed credentials, and each authentication message will include a nonce and timestamp to prevent replay or tampering. Additionally, the AI-based anomaly detection layer will monitor flight behavior for deviations that may indicate control interception, triggering re-authentication and blockchain verification.

Comment 5: It is highly suggested to improve the quality of some figures. For example, Fig. 4, 5, and 6. It seems like they were pasted from another book or article. These figures can be easily improved for changing their quality.

 

Response: We sincerely thank the reviewer for this observation. We acknowledge that the visual quality of Figures 4, 5, and 6 in the original submission did not fully meet publication standards due to the way they were exported. In the revised manuscript, these figures have been completely re-created using high-resolution vector graphics and direct outputs from our experimental environment, rather than screenshots or pasted images. The text, symbols, and line work have been redrawn to ensure sharpness, consistency, and clarity in both digital and print formats. The updated figures now maintain full readability when zoomed in and are entirely original to this work.

 

Comment 6: Some practical applications on the drone systems should be investigated and discussed in this paper. a) "Performance Analysis and Optimization Design of AAV-Assisted Vehicle Platooning in NOMA-Enhanced Internet of Vehicles," in IEEE Transactions on Intelligent Transportation Systems; b) "A AAV Swarm Authentication and Key Agreement Scheme Based on Latin Square Design," in IEEE Wireless Communications Letters (the author is free to cite recommendations).

 

Response: We thank the reviewer for pointing us toward these valuable contributions. In the revised manuscript, we added a dedicated paragraph in the Related Works and Discussion sections to address practical applications of autonomous aerial vehicles (AAVs)/unmanned aerial vehicles (UAVs) in drone systems and positioned our work relative to the recommended literature. This is as follows.

 

Comment 6: Before the final document version, the authors should do a grammatical review of the paper. For example, the title of Figure 8 “based on the VisioDect [?]”.

 

Response: The authors carried out thorough grammatical review of the revised manuscript, from the abstract to the conclusion section. Also, the caption of the former Figure 8 (now Figure 11) was corrected. For instance,

Figure 11. Samples of Real-Time Drone detection and Attached Objects recognition and elicitation by DRONET [12] and ALIEN [5] based on the VisioDect [32] dataset showing high prediction precision and sensitivity on different climatic scenarios necessary for the iBANDA authentication and eventual neutralization.

Also, different sections of the revised manuscript were updated. E.g.

To provide smart and safe aerial logistics operations, legacy-based DDS (empowered by AI and IoT) takes between 100 ms to 1700 ms to authenticate the validity of a drone within its network range before initiating an appropriate counter-response [13]. Albeit with several security loopholes. Safe and secure drone-based logistics operations could be revolutionized by integrating blockchain technology into the DDS networks, which provide a secure, traceable, decentralized, and tamperproof platform for sending, storing, and verifying the source of sensitive data [14]. However, this additional blockchain layer in the DDS network can result in increased DDS complexity and response delay while improving authentication validity. Due to the necessity of balancing security, speed, and scalability—a problem that has persisted in blockchain-assisted designs—the application of blockchain technology in DDS is still in its infancy

Overall, the authors wish to thank the reviewer for the insightful reviews and constructive criticisms on the submitted manuscript which have helped the authors to improve the quality of the revised manuscript in terms of content and clarity.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Methodology – Soundness and Rigor

Strengths of the Manuscript

• Novel Dual-Layer Authentication Framework:

The manuscript presents a new concept of iBANDA, which combines the AI assisted drone and payload recognition with an authentication using ’blockchain’ technology. Traditional ID methods deal only with the drone itself and not the payload, posing a serious security risk for air freight via drones — iBANDA is unique in that sense; it authenticates both.

• Edge-Lightweight Trainer: 

A real-time AI model that reduces the need for large memory-intensive pre-trained models using transfer learning.

The authors’ approach is practical and resource-aware onboard drone detection using a quantized YOLOv5n model for the NVIDIA Jetson Nano. The model attains a mAP of 99.5% and inference time of 0.021s, outperforming several other deep learning baselines with minimal compute overhead (16.10 GFLOPS), which is quite commendable.

• Efficient Consensus Mechanism (Snowball PoS):

The Avalanche Snowball-based Proof of Stake (PoS) consensus mechanism provides a low-latency, scalable and energy-efficient alternative to traditional blockchain schemes (PoW, PoA). The recurrent sub-sampled voting of the consensus algorithm prevents Sybil and DDoS attacks and is a key element to achieve resilience at decentralizing drone environments.

Suggestions for Improvement (Technical and Methodological)

• Incomplete Parameterization in Mathematical Formulation

The paper introduces a structured authentication formulation (e.g. Equations (1) to (6)), but leaves key variables and thresholds as symbolic references rather than instantiating them. For example:

The Snowball consensus leaves α (quorum size) and β (confidence threshold) undefined.

There is no advice given on how to balance system responsiveness vs. fault tolerance.

Recommendation:

Listing some intuitive parameter ranges (α = 5, β= 3), or establishing a trade-off between the system responsiveness and consensus robustness (against Sybil attacks or collusion) with some heuristic table.

 

• Results Without Convergence and Security Guarantees for the Snowball Algorithm

No convergence probability and adversarial resistance of the Snowball consensus algorithm, based on provided no quantitative statements.

Recommendation:

Add a theoretical explanation or bounded (e.g., number of expected rounds to consensus, Byzantine resistance, etc.). in Avalanche/Snowball literature) declaration Perform pseudo-adversarial convergence to show robustness

 

• No Formal System Modeling or State Transition Diagram

While the DDS pipeline is illustrated using a logic diagram, it lacks a formal state transition model (e.g., FSM or Petri Net) that captures its transitions in normal and anomalous drone behaviors.

Recommendation:

Supplement the system design section with a discrete state transition diagram showing:

• Pre-authentication → Authentication → Neutralization decision states

• Transitions based on data mismatch, missing credentials, or consensus failure

 

• Need for a Comparative Ablation Study

Although the comparison is extensive, the paper lacks a controlled ablation study separating out the different effects of each iBANDA component.

Blockchain inclusion vs. legacy DDS

AI-only model vs. AI + BC

 

Snowball vs. PoA/PoW consensus

 

Recommendation:

Show at least one table/graph of ablation here to quantify the performance delta.

We evaluated latency, TPS, and false positive rate among all system configurations.

• Limited Scalability at the Network Layer

Whereas TPS metrics are often reported, it is unclear what the horizontal scalability of the blockchain network will be (i.e., how many drones per validator node? Performance degradation due to node churn? is not evaluated.

 

Recommendation:

Simulate network scaling effects on:

• Latency (µt1 and µt2) with more than 50 concurrent drones

 

• The effect of the size of a validator pool on transaction confirmation time.

 

Additional Notes

Formatting Irregularities:

Equations are sometimes in-line without proper LaTeX styling ( e.g. D↑ =... ) 

The authors should consider formatting the equations properly in MathJax or Latex style for readability.

Citation Style:

A few references (e.g., " [6? ], "[?]" ) look odd or malformed, which probably means missing citation placeholders or a lack of BibTeX entries. This needs correction before publication.

 

 

Comments on the Quality of English Language

The language structure of the manuscript is quite  readable and accademically  structured in the English language setting, but the work can benefit from sound professional touch. I encourage a round of language polish be done before the final closing.

Author Response

Comment 1:

• Incomplete Parameterization in Mathematical Formulation

The paper introduces a structured authentication formulation (e.g. Equations (1) to (6)), but leaves key variables and thresholds as symbolic references rather than instantiating them. For example:

The Snowball consensus leaves α (quorum size) and β (confidence threshold) undefined. There is no advice given on how to balance system responsiveness vs. fault tolerance.

Recommendation:

Listing some intuitive parameter ranges (α = 5, β= 3), or establishing a trade-off between the system responsiveness and consensus robustness (against Sybil attacks or collusion) with some heuristic table.

 

Response: We appreciate the reviewer’s observation regarding the symbolic treatment of α (quorum size) and β (confidence threshold) in our Snowball consensus formulation, and the request for guidance on parameter selection. In the revised manuscript, we have now included intuitive parameter ranges and a rationale for their selection, supported by experimental validation. To highlight the need to instantiate and contextualize the key Snowball consensus parameters, we have: explicitly instantiated parameters; added practical guidance; and included a Rule-of-thumb formula. These additions make the formulation operationally actionable, allowing practitioners to configure α and β to match mission priorities while preserving the security guarantees of the Snowball-based PoS consensus. Section 3.3.1 captures these details in the revised manuscript as shown.

 

Comment 2: No convergence probability and adversarial resistance of the Snowball consensus algorithm, based on provided no quantitative statements. Recommendation: Add a theoretical explanation or bounded (e.g., number of expected rounds to consensus, Byzantine resistance, etc.) declaration. Perform pseudo-adversarial convergence to show robustness.

 

Response: We thank the reviewer for this important observation. In the revised manuscript, we have augmented Section 3.2.2 with both theoretical bounds from Avalanche/Snowball literature and our empirical convergence results under pseudo-adversarial conditions.

 

Comment 3: No Formal System Modeling or State Transition Diagram. While the DDS pipeline is illustrated using a logic diagram, it lacks a formal state transition model (e.g., FSM or Petri Net) that captures its transitions in normal and anomalous drone behaviors.

Response: We updated the manuscript by including the FSM and Petri Net as suggested by the reviewer. The details of these are found in Section 3 of the revised manuscript.

 

Comment 4:

• Need for a Comparative Ablation Study

Although the comparison is extensive, the paper lacks a controlled ablation study separating out the different effects of each iBANDA component.

Blockchain inclusion vs. legacy DDS

AI-only model vs. AI + BC

Snowball vs. PoA/PoW consensus

Recommendation:

Show at least one table/graph of ablation here to quantify the performance delta.

We evaluated latency, TPS, and false positive rate among all system configurations.

 

Response: We thank the reviewer for this insightful comment. In the revised manuscript, we have conducted a controlled ablation study to isolate the performance contribution of each iBANDA component. Specifically, we evaluated:

1. Blockchain Inclusion vs. Legacy DDS – to measure the trade-off between security and performance.
2. AI-only Model vs. AI + Blockchain – to determine the improvement in detection accuracy and false positive rate.
3. Consensus Mechanisms (Snowball vs. PoA/PoW) – to compare the transaction throughput and latency impact of different consensus strategies. Section 4.4 presents these details.

Comment 5: Formatting Irregularities:

Equations are sometimes in-line without proper LaTeX styling ( e.g. D↑ =... ) 

The authors should consider formatting the equations properly in MathJax or Latex style for readability.

 

Response: We sincerely thank the reviewer for pointing out the inconsistency in the equation formatting (e.g., use of D↑ = ... without proper LaTeX styling). We agree that proper formatting using MathJax/LaTeX will greatly improve readability and presentation. For instance,

 

Comment 6: Citation Style:

A few references (e.g., " [6? ], "[?]" ) look odd or malformed, which probably means missing citation placeholders or a lack of BibTeX entries. This needs correction before publication.

 

Response: We updated the manuscript by fixing these omitted citations.

 

Overall, the authors wish to thank the reviewer for the insightful reviews and constructive criticisms on the submitted manuscript which have helped the authors to improve the quality of the revised manuscript in terms of content and clarity.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper makes an original and relevant contribution to the field of drone-based logistics security. The proposed iBANDA system, a novel AI and blockchain-assisted defense framework, addresses the critical problem of drone identity spoofing and unauthorized payload transport. With minor revisions, I consider it will be suitable for publication.

While the testbed is adequately described overall, the information provided to reproduce the experiments (e.g., those associated with Table 5 - AI Model Detection Performance) is not sufficient for exact replication. I would appreciate additional details.

Some adjustments and corrections are required before publication, such as the following:

• In the Abstract, “YOLOv5s” should be briefly introduced.
• In the first paragraph on page 2, avoid using “etc.” in the middle of a sentence. For instance, replace it with “among other concerns.”
• All tables, figures, algorithms, and theorems should appear only after being cited in the main text. For example, Table 1 appears on page 2 but it is not mentioned until page 4.
• Table 2 is never referenced in the text.
• The formatting of the text immediately following Table 2 requires correction.
• The term “Lidar” should be introduced as “Light Detection and Ranging (LIDAR)” the first time it appears (page 4), and “LIDAR” should be used consistently thereafter.
• On page 5, before “Theorem 1,” the reference should read “as specified in Theorem 1” instead of “as specified in Theorem (2).”
• On page 6, the number of Equation (6) should be right-aligned.
• On page 7, the presentation of the considerations list should be improved. For instance, remove the word “That” from the first two items and apply consistent and appropriate end punctuation.
• On page 8, Table 5 is mentioned before Tables 3 and 4 are introduced or discussed.
• On page 9, the reference “in Theorem ??T0.1)” should be corrected to “in Theorem 2.”
• On page 10, the sentence “If or more validators” should be corrected to “if α or more validators.”
• In the last paragraph of page 12, “proof of work (PoA)” should be corrected to “proof of authority (PoA).”
• On page 13, when referring to Table 5, the evaluation metrics used should be further explained.
• Figure 8 requires a more detailed description in the text that clearly describes it.
• The dataset named “VisioDect [?]” in the title of Figure 8 must be properly cited.
• In Figure 9, the horizontal axis for the number of transactions should begin at 8.
• On page 16, the phrase “Table ?? shows” should be corrected to “Table 6 shows”.

Author Response

Comment 1: While the testbed is adequately described overall, the information provided to reproduce the experiments (e.g., those associated with Table 5 - AI Model Detection Performance) is not sufficient for exact replication. I would appreciate additional details.

 

Response: We thank the reviewer for this important observation regarding the reproducibility of the AI testbed experiments, particularly those associated with Table 5 (AI Model Detection Performance). We agree that additional implementation details will strengthen the replicability of our work. Notwithstanding, the authors wish to restate that the full details of the AI model implementation can be found in the DRONET [12] and ALIEN [5] of our previous work, which the readers’ attention has been drawn to (See Section 3: IBANDA Framework).

That being said, we have expanded the testbed description to provide the following additional information:

1. Dataset and Preprocessing:

• The VisioDect dataset [31] was used, containing annotated UAV and payload images under varying climatic conditions.
• Images were resized to 128×128 pixels prior to training, and data augmentation (rotation ±15°, horizontal flipping, contrast normalization) was applied to enhance robustness.

1. Model Configuration:

• The AI module was implemented using YOLOv5s (small) with PANET for feature aggregation.
• The network was quantized to ONNX for deployment and pruned by removing non-critical convolutional layers and filters to minimize complexity.
• Training was conducted with batch size = 32, learning rate = 1e-3, and Adam optimizer with cosine annealing scheduler.

1. Hardware/Software Environment:

• Training was performed on an NVIDIA RTX 3090 GPU (24 GB VRAM) with PyTorch 1.12 and CUDA 11.6.
• The edge deployment used an NVIDIA Jetson Nano (4 GB RAM, Quad-core ARM Cortex-A57), validating low-compute performance.

1. Evaluation Metrics:

• Mean Average Precision (mAP), Recall (R), F1-score, Inference Time (fps), Relative Loss, and GFLOPS were recorded.
• Reported results in Table 5 represent averages over five independent runs to account for variance.

These additions provide the necessary information for an independent researcher to replicate our results with high fidelity.

 

Minor Comments:

Some adjustments and corrections are required before publication, such as the following:

• In the Abstract, “YOLOv5s” should be briefly introduced.

 

Response: We recast the entire abstract section and defined all terms accordingly. Moreso, we include a Table containing list of acronyms used in the revised manuscript.

 

• In the first paragraph on page 2, avoid using “etc.” in the middle of a sentence. For instance, replace it with “among other concerns.”

Response: We deleted the “etc.” in the middle of the sentence.

 

 

 

• All tables, figures, algorithms, and theorems should appear only after being cited in the main text. For example, Table 1 appears on page 2 but it is not mentioned until page 4.

Response: We have adjusted this accordingly and maintained this golden rule across the revised manuscript.

 

• Table 2 is never referenced in the text.

Response: Former Table 2 (now Table 14) has been referenced in the revised manuscript and has been moved to the last page of the manuscript (before reference section) in line with MDPI style.

 

• The formatting of the text immediately following Table 2 requires correction.

Response: This has been fixed as seen.

“Modeling a DDS to address comprehensive and dynamic drone operation challenges, often referred to as “NP-problems,"…”

 

• The term “Lidar” should be introduced as “Light Detection and Ranging (LIDAR)” the first time it appears (page 4), and “LIDAR” should be used consistently thereafter.

Response: This has been corrected.

“It uniquely integrates AI-based visual and Light Detection and Ranging (LIDAR) sensing with lightweight blockchain authentication,…”

 

• On page 5, before “Theorem 1,” the reference should read “as specified in Theorem 1” instead of “as specified in Theorem (2).”

Response: We have corrected this.

 

 

• On page 6, the number of Equation (6) should be right-aligned.

Response: All equations in this work are centralized to maintain uniformity.

 

• On page 7, the presentation of the considerations list should be improved. For instance, remove the word “That” from the first two items and apply consistent and appropriate end punctuation.

Response: We have updated these statements as seen.

• The drone detection and attached object identification algorithm is intact and controlled by an AI object detection model, as seen in the works of [12];
• The scenario-specific neutralization response framework is intact as proposed in our previous work, the ALIEN model [5];
• There is an overseeing organization that governs airspace activities, acting as the

primary authority for drone operations and security [4];

• Each drone employed for smart mobility is licensed to operate and properly registered in the network, classified either as a freight operator or a hobby drone user with the relevant regulatory authority [31]; and
• The participating registered drone users and logistics operators serve as validators for the mining process in the blockchain network.

 

• On page 8, Table 5 is mentioned before Tables 3 and 4 are introduced or discussed.

Response: The statement mentioning the Table 5 has been deleted in the revised manuscript.

 

• On page 9, the reference “in Theorem ??T0.1)” should be corrected to “in Theorem 2.”

Response: We have corrected this statement in the revised manuscript.

 

• On page 10, the sentence “If or more validators” should be corrected to “if α or more validators.”

Response: We have corrected the statement in the revised manuscript as shown.

“As seen in Algorithm 1, the initial transaction preference is set as legit − drone.

Until a transaction has been decided, the sample size (k) is queried, and each validator is requested to make a preference. If α or more validators give the same response, the…”

 

• In the last paragraph of page 12, “proof of work (PoA)” should be corrected to “proof of authority (PoA).”

Response: This has been fixed accordingly.

 

• On page 13, when referring to Table 5, the evaluation metrics used should be further explained.

Response: The authors have broadened the discussion of Table 5 results by explaining the meaning and implication of each of the metrics.

 

 

• The dataset named “VisioDect [?]” in the title of Figure 8 must be properly cited.

Response: This has been corrected as seen.

“Figure 8. Samples of Real-Time Drone detection and Attached Objects recognition and elicitation by DRONET [11] and ALIEN [5] based on the VisioDect [28] dataset showing high prediction precision and sensitivity on different climatic scenarios necessary for the iBANDA authentication and eventual neutralization.”

 

• In Figure 9, the horizontal axis for the number of transactions should begin at 8.

Response: The figure has been redrawn to capture this detail.

 

 

• On page 16, the phrase “Table ?? shows” should be corrected to “Table 6 shows”.

Response: The authors have fixed this issue in the revised manuscript.

As seen in Table 7, it takes an average Dapp time of 2293 ms to complete the

 

Overall, the authors wish to thank the reviewer for the insightful reviews and constructive criticisms on the submitted manuscript which have helped the authors to improve the quality of the revised manuscript in terms of content and clarity.

 

 

 

 

 

 

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Accept in present form.