Authenticated Transfer of Misbehaviour Evidence Across Pseudonym Changes in Vehicular Networks
Mavis Serwaa Boateng, Adasa Nkrumah Kofi Frimpong, Eldad Antwi -Bekoe, Salam Fuseini
Abstract
Vehicles in a pseudonymous vehicular network change identifiers to protect the location privacy of drivers. The same change erases what a misbehaviour detector has learned about a vehicle, so a misbehaving vehicle can shed its record by changing pseudonym. Earlier work restores the link by re-identifying the vehicle from its movement. That link is a statistical guess, and the method is itself a tracking attack. We present IV-POC, a protocol that moves the misbehaviour evidence record of a vehicle from its old pseudonym to its new one as a signed, encrypted, single-use transfer. The evidence owner freezes the record, the certificate authority and a privacy guardian must both decrypt it, three of five authorities approve it, and the receiving side installs it exactly once. We evaluate the artefact on the public VeReMi Extension dataset with a three-arm design: evidence reset as the lower bound, perfect continuity as the upper bound, and IV-POC between them, with zero-evidence controls. In a verified run over 11 roadside placements, the state installed after each of 454 transfers matched the perfect-continuity reference exactly across 26,520 replayed records, and the implementation rejected forged, replayed and under-approved transfers in all 5,720 authorisation checks. The transfer recovered the whole gap between reset and perfect continuity, and with corroborating witnesses it revoked a scheduled case a median 9 s earlier than a reset, which missed two cases altogether. Continuity also halved the honest false-flag rate under a fixed threshold, and a learned threshold removed that penalty. A transfer carried a median 9.2 kB and 2.1 ms of processing. These results come from one trace excerpt of one attack family; the full campaign is in progress.
Review reports
xPeerd (AI review)
AI Review (xPeerd — DAReview)
Summary
This manuscript addresses a genuine and underexplored systems problem in V2X security: pseudonym changes protect privacy but can erase accumulated misbehaviour evidence at exactly the moment when a detector is close to action. The paper proposes IV-POC, a signed, encrypted, approval-based transfer protocol that moves the detector state from an old pseudonym to a new one, and it evaluates the mechanism on the VeReMi Extension dataset using three arms: reset, authenticated transfer, and perfect continuity. The paper’s strongest qualities are its clear problem framing, explicit articulation of assumptions, reproducibility-oriented implementation mindset, and a bounded evaluation design using lower- and upper-bound comparators. The reported verified run shows exact equality between transferred and perfect-continuity state across 454 transfers and successful rejection of malformed or unauthorised attempts in all reported checks. The paper also reports practical overhead figures and explicitly acknowledges important threats to validity.
Its main weakness is that the empirical evidence remains too narrow for strong performance claims. The manuscript itself states that “Every number in Sections 4.2 to 4.5 comes from the single verified job described below; replace them with the campaign values before submission,” and later concedes that “the full 16-folder campaign is still running on the authors' host.” This substantially limits the generality, stability, and publishable maturity of the evaluation in its present form.
Potential Major Revisions
- The empirical validation is currently too narrow to support strong general conclusions.
The central limitation of the manuscript is that the reported quantitative evidence is explicitly based on a single verified job rather than the full campaign. In Section 4.1, the authors state: “The full 16-folder campaign is still running on the authors' host. Every number in Sections 4.2 to 4.5 comes from the single verified job described below; replace them with the campaign values before submission.” This is reinforced again in the abstract, which ends with the caution: “These results come from one trace excerpt of one attack family; the full campaign is in progress.”
This directly weakens the manuscript’s claims about effectiveness, revocation benefit, false-flag reduction, and operational cost. A result derived from one “13-receiver sample of the constant-position-offset family” cannot yet establish robustness across attack families, densities, RSU placements, or traffic periods. The manuscript should not merely mention this as a limitation; it should either complete the full campaign before submission or substantially narrow the claims in the title, abstract, results, and conclusion. At present, the wording risks overstating maturity relative to the presented evidence.
- The protocol’s correctness is demonstrated at the implementation level, but not analysed formally.
The manuscript is appropriately candid here, but this remains a major gap. Section 4.3 states: “These are tests of an implementation. They are not a proof of the protocol, and a formal analysis in the manner of [23], [24] remains future work.” For a security-sensitive transfer protocol involving decryption shares, approval thresholds, key possession proofs, and single-use guarantees, implementation-level rejection tests are useful but insufficient as the main assurance argument.
This matters because the paper’s contribution is not just empirical continuity, but a cryptographic and systems mechanism with claimed security properties. Section 3.4 lists strong enforced properties such as “Exclusivity,” “Single use,” “No summation,” “Operating-point binding,” and “Age preservation.” Yet these are currently asserted as implementation properties rather than proven under an adversarial model. A major revision should include either a formal protocol analysis or, at minimum, a structured threat model and argument mapping each claimed property to concrete attack surfaces, trust assumptions, and failure modes.
- The trust and deployment assumptions are strong enough to limit practical realism, and the paper does not fully convert that into a deployment analysis.
Section 3.1 explicitly relies on assumptions A1–A6. Several are especially restrictive: “one runtime owns all records,” “the gateway that combines their decryptions is trusted and learns the link,” “at least three of the five approval authorities are available and honest,” “the transfer travels over a protected infrastructure link,” and “the vehicle signs with its old and new keys.”
The manuscript does acknowledge some of these constraints. For example, Section 4.6 states: “One runtime owns all records, so the results say nothing about handover between RSUs (A1),” and concludes that “An operator that cannot grant that trust should not deploy this design as it stands.” This honesty is commendable, but the current paper still reads more like a deployment-ready architecture than a constrained proof-of-concept. The revision should more sharply distinguish what has been shown from what remains speculative. In particular, the paper should discuss:
- how often the required joint availability of CA, PG, gateway, and 3/5 authorities is realistic in live C-ITS backends;
- what happens under partial failure, delay, or desynchronisation;
- how security and accountability change if the gateway is semi-honest rather than fully trusted;
- whether deployment requires policy changes to enforce the manuscript’s statement that “a deployment must refuse a new pseudonym until a transfer receipt exists.”
- The privacy claim is carefully bounded, but still under-analysed relative to the significance of the contribution.
The paper positions itself against inference-based tracking by arguing that its approach restricts linkability to a specific object and moment. Section 2.1/2.4 states: “IV-POC uses explicit linkability restricted to one object, the evidence record, and to one moment, the change.” However, Section 3.1 also states: “The packet contains signatures by both pseudonym keys, so a party that reads the packet and knows the public keys could link them. We claim unlinkability against observers of the radio channel only.”
This is an important limitation, not a minor caveat. The contribution is partly justified as avoiding the privacy costs of re-identification, yet the manuscript does not provide a systematic privacy analysis of infrastructure observers, internal entities, retention risks, metadata exposure, or governance implications of the gateway learning the link. Section 4.5 also notes that the protocol “does not hide the link from the gateway (A2), does not protect a packet read on the infrastructure link (A4), and stores the list of counted observation identifiers in the record.” A major revision should therefore include a dedicated privacy analysis section that makes the privacy tradeoff explicit rather than implicit.
- The evaluation design is thoughtful, but the effect-size claims remain contingent on scheduling choices and a favourable transfer model.
The experimental structure is good: reset is a lower bound, perfect continuity an upper bound, and a zero-evidence schedule a negative control. But the observed gains depend heavily on design choices that favour isolating the target effect. Section 4.6 acknowledges that “Each selected sender changes pseudonym once, at a point chosen where evidence exists; that isolates the effect we want to measure and overstates how often a change coincides with pending evidence in ordinary traffic.” It also states: “Transfers always complete and take no simulated time, so we do not model a lost packet, an unavailable authority or a vehicle that refuses to sign (A5).”
This means the paper currently demonstrates mechanism value under idealised eligibility and completion conditions, rather than likely field performance. The revision should add sensitivity analysis around failed transfers, delayed transfers, repeated pseudonym changes, sparse coverage, and varying evidence-age distributions. Without that, the result that “G = 1” is correct within the defined setup, but not yet persuasive as a real-world performance conclusion.
- The manuscript’s novelty claim is plausible, but the comparative evaluation against the closest alternative is incomplete.
The paper identifies V-TRACE as “the closest work” and contrasts explicit transfer with inference-based relinking. It also states that the evaluation currently lacks “a comparison with inference-based re-identification on the same traces” and lists that as future work near the end of the manuscript. Because the paper’s central thesis is that explicit authenticated continuity is preferable to probabilistic tracking, the absence of a direct empirical comparison leaves the practical advantage only partially demonstrated. A major revision should compare IV-POC against the closest inference baseline under the same traces, metrics, and deployment assumptions, including privacy and latency tradeoffs.
- Some claimed operational efficiency is promising, but its deployment implications remain underdeveloped.
Section 4.5 reports small median processing costs and packet sizes in the verified run, but it also acknowledges two important cautions: “a real record costs about thirteen times more per observation than the synthetic benchmark” and “a packet of several kilobytes is far larger than a safety message of a few hundred bytes.” The manuscript further states that “a long-lived record would need a compact representation of counted observations before deployment.”
This is a major issue because storage growth and transmission size are not peripheral details here; they directly affect deployability. If the evidence record grows with age, then the architecture may be practical only under constrained detector-state designs or short evidence windows. The manuscript should analyse worst-case growth, backend throughput, storage pressure, and record compression strategies rather than leaving compact representation to future work.
- Some template or placeholder content remains, which raises integrity and readiness concerns.
Near the end of the manuscript, the “Funding” line reads: “None, or state the funder and grant number.” Under “Data and Code Availability,” the paper says the artefacts are available at “repository or archive link.” These are clear submission placeholders. Their presence suggests the manuscript is not yet in final review-ready form. This is not misconduct, but it does indicate incomplete editorial preparation and creates uncertainty about the availability of the reproducibility materials that the paper relies on as part of its contribution.
Potential Minor Revisions
- Incomplete metadata and placeholders
- “None, or state the funder and grant number.” in the Funding section is an unresolved placeholder and should be replaced with a definitive statement.
- “repository or archive link” in Data and Code Availability is also a placeholder and should be replaced with a persistent, public identifier if the materials are available.
- The sentence in Section 4.5, “Re-run bench_transfer.py on the authors' host and replace them,” reads like an internal drafting instruction rather than polished manuscript prose.
- Overly self-referential drafting language Several passages signal work-in-progress status too directly for a finished paper and should be revised for publication tone:
- “replace them with the campaign values before submission” in Section 4.1
- “the full campaign is in progress” in the abstract These are transparent, but in a submitted paper they read as internal process notes rather than stable scientific reporting.
- Clarity and wording refinements
- The sentence “Our central question is not whether the scheme works. It is under what conditions it works, and whether those conditions are realistic” is conceptually strong, but it would read more precisely if tied immediately to explicit evaluative criteria.
- “The transfer recovered the whole gap between reset and perfect continuity” is mathematically correct within the manuscript’s defined design, but rhetorically strong enough to invite overinterpretation beyond the single-job evaluation.
- “Continuity also halved the honest false-flag rate under a fixed threshold” is impactful, but the manuscript should repeatedly remind readers that this is trace- and threshold-specific.
- Figure and table consistency
- The manuscript refers to “Fig. 3” and later “Fig. 6” in the overhead discussion; the numbering appears internally consistent in the excerpts, but the final layout should be checked carefully because the narrative depends heavily on figure cross-reference precision.
- Table 2 and Table 5 are useful, but the paper would benefit from a compact notation table summarising symbols such as G and R before they are used in prose-heavy interpretation.
-
Stylistic compression The manuscript is generally clear and disciplined, but some sentences are longer than necessary in methods sections and assumptions sections. For example, A4 compresses several privacy conditions into one paragraph and would benefit from being split into claim, threat surface, and consequence.
-
AI content analysis I do not find strong evidence here of generic AI-style fabrication or meaningless verbosity. The manuscript is technically coherent, domain-specific, and internally cross-referenced in a way that is more consistent with careful human drafting than with low-effort generated text. Its strongest signs of authenticity are the explicit assumptions, concrete protocol steps, nuanced threats-to-validity discussion, and willingness to report limitations against its own claims.
Estimated percentage of AI-generated content: low to moderate uncertainty; no reliable document-internal basis for a numerical attribution. I therefore do not consider a percentage estimate methodologically defensible from the manuscript text alone.
Highlighted AI-detected sections:
- No section stands out as obviously machine-generated.
- The placeholder lines in Funding and Data/Code Availability are editorial incompletions, not evidence of AI generation.
Assessed epistemic impact:
- Even if limited generative assistance had been used in drafting, the epistemic quality of the manuscript would depend mainly on the truthfulness of the protocol description, reproducibility assets, and empirical reporting.
- The current epistemic risk comes not from likely AI generation, but from the narrowness of the presented evaluation relative to the strength of some framing claims.
- Fraud / paper-mill screening
I do not see strong paper-mill traits such as incoherent citations, meaningless terminology chains, impossible methodology, or fabricated-looking quantitative structure. On the contrary, the manuscript explicitly discloses limitations such as simulated sensors, synthetic RSUs, single-run reporting, and unmodelled failures.
Warning: the paper is not obviously fraudulent, but it is not yet fully publication-ready because some artefact links and evaluation components remain incomplete.
Recommendations
-
Complete the full 16-folder campaign before submission and rewrite the abstract, results, and conclusion to reflect aggregated results across attack families, traffic periods, and RSU placements. If this is not feasible, reposition the paper explicitly as a proof-of-concept or design-and-validation study rather than a broadly validated performance paper.
-
Add a dedicated security and privacy analysis section. This should enumerate adversaries, what each party can observe, what linkability is exposed, which guarantees rely on trust rather than cryptography, and how the gateway’s privileged role affects the privacy argument.
-
Strengthen the assurance argument by adding either formal protocol analysis or a more systematic attack analysis. At minimum, map each claimed property from Section 3.4 to explicit attacks, assumptions, and empirical tests, and identify what remains unproven.
-
Evaluate non-ideal operating conditions: failed or delayed transfers, unavailable authorities, sparse infrastructure, repeated pseudonym changes, and vehicle refusal to sign. These are already acknowledged as limitations, so converting them into experiments would substantially improve credibility.
-
Add a direct comparison against the closest inference-based continuity approach on the same traces if possible. Since the manuscript’s rationale partly rests on avoiding tracking-style re-identification, an empirical side-by-side comparison would materially strengthen the novelty and practical relevance claims.
-
Expand deployment realism discussion. In particular, explain backend coordination requirements, operational latency budgets, governance assumptions for CA/PG/gateway separation, and how the “refuse a new pseudonym until a transfer receipt exists” requirement would fit actual V2X credential management workflows.
-
Finalise all editorial placeholders and make the reproducibility package concretely accessible with stable identifiers, versioned artefacts, and exact instructions sufficient for independent reruns.
Overall, I judge the manuscript to be technically interesting, coherent, and potentially publishable after substantial revision. Its conceptual contribution is credible, but the present evidence base is too limited, and the security/privacy/deployment argument remains too assumption-dependent, for acceptance in its current form.
XSci (AI review)
Summary
In a pseudonymous vehicular network a change of identifier erases what a misbehaviour detector has accumulated about a vehicle, so a vehicle can shed its record by rotating pseudonyms. The authors propose IV-POC, which moves the detector's evidence record to the new pseudonym as a signed, encrypted, single-use transfer: the source owner freezes and snapshots the record and splits it into shares encrypted to a certificate authority and a privacy guardian, signatures under both pseudonym keys and a three-of-five approval certificate are collected, and a gateway recombines the shares so the destination installs the state exactly once. The scheme is evaluated on a 13-receiver excerpt of the constant-position-offset family of VeReMi Extension under 11 roadside-unit placements, with reset, transfer and perfect-continuity arms and zero-evidence controls. Across 454 transfers the installed state matched the reference in all 26,520 replayed records, 5,720 authorisation checks rejected every malformed attempt, and a transfer cost a median 9.2 kB and 2.07 ms.
Contribution and Significance
The contribution that matters is not the cryptography, which composes standard primitives (ECIES over P-256, exclusive-or sharing, ECDSA, threshold approval) and which the authors do not claim. It is, first, the framing: evidence continuity is distinct both from back-end revocation, which acts only after a decision has been taken, and from inference-based re-identification, which guesses the link. Second, and more durable, it is the measurement design of Fig. 2 — reset and perfect-continuity bounds bracketing the treatment, a schedule frozen by a held-out eligibility replay before any arm produces an outcome, and a zero-evidence negative control. That instrument can score any future continuity mechanism on the same scale, and is what other groups are most likely to reuse.
Relative to the closest prior work the novelty is narrower than Section 1 and Table 1 suggest. V-TRACE [8] and linkage values [1] are correctly positioned, but authority-mediated transfer of accumulated per-pseudonym state across an identifier change is not new. Miranda and Rodrigues ("Reputation in anonymous vehicular networks," International Journal of Autonomous and Adaptive Communications Systems, vol. 3, no. 2, pp. 178–197, 2010) transfer accumulated reputation to a new pseudonym using blind signatures, so that no party learns the link, and block the acquisition of multiple personalities; CoPRA (N. Bißmeyer, J. Petit and K. M. Bayarou, Proc. 10th IFIP/IEEE WONS, Banff, 2013, pp. 9–16) releases a pseudonym link only after infrastructure validates the stated reason, which is the pattern of the authors' approval request. Neither is in the reference list. Genuinely new are the object transferred — a full detector state, including counted observations and the operating-point digest — the no-summation and operating-point-binding properties, and the evaluation design.
Strengths
- The exactness result is clean and correctly scoped: all 26,520 replayed records agreed between the transfer and perfect-continuity arms (Section 4.2), and the authors state plainly that these are implementation tests, not a proof.
- The control design is unusually careful: 396 of the 454 transfers are zero-evidence controls, and all three arms share messages, keys, thresholds and transition points (Fig. 2, Table 2).
- The reproducibility machinery — 66 protocol checks per job, three independent signed replays, a separate verifier and SHA-256 receipts binding every output (Section 4.1) — exceeds normal practice in this field.
- Assumptions A1–A6 and Section 4.6 are candid about zero transfer delay, synthetic roadside units, two attacker vehicles, no inter-RSU handover and the trust placed in the gateway.
- The incidental finding that dropping the record inflates honest flagging from 3.60 to 6.83 per cent under the fixed threshold, and that the conformal threshold absorbs this (Section 4.3), is a useful observation about detector design.
Recommendation
Reject for Vehicular Communications. Two of the ways a manuscript falls short at this journal apply: it is too incomplete to assess, since Section 4.1 says its numbers must be replaced before submission and the availability section is empty; and the evidence does not convincingly support the central claim about the value of continuity, which is definitional in its headline form and rests on two revocation failures of one vehicle in its empirical form. Neither is a flaw in the design, and a version built on the completed campaign, with a privacy analysis and the missing prior work engaged, would be worth submitting — but that is more than one revision round of work.
Reject for Ad Hoc Networks. The same concerns apply, and that readership would also expect a formal analysis of the protocol rather than tests alone.
Major Revision for Wireless Personal Communications. The protocol, the exactness result and the overhead measurements would suffice there once the placeholder text is removed, the claims are restated at the strength the data support, and the missing prior work is cited.
Major Concerns
-
The manuscript is an unfinished draft by its own account (Sections 4.1, 4.5; Conflict of Interest; Data and Code Availability). Section 4.1 states: "The full 16-folder campaign is still running on the authors' host. Every number in Sections 4.2 to 4.5 comes from the single verified job described below; replace them with the campaign values before submission." Section 4.5 repeats this for the benchmark: "Re-run bench transfer.py on the authors' host and replace them." The Conflict of Interest section reads "None, or state the funder and grant number"; Data and Code Availability is empty; and the availability sentence, misplaced under References, ends "available at repository or archive link." A referee cannot assess numbers the authors themselves call provisional, and the open implementation claimed as contribution 2 cannot be checked.
-
The headline detection result follows from the construction, and the one quantity that does not rests on a single vehicle (Section 4.4, eq. (2), Table 4). The transferred state is bit-identical to the reference (Section 4.2), the transfer takes zero simulated time (Table 2), and no transfer fails ("Transfers always complete and take no simulated time," Section 4.6), so the transfer arm cannot differ from the perfect-continuity arm and G = 1 could not have come out otherwise; the authors concede as much. The informative quantity is the denominator, 0.091, which is two reset failures that Section 4.4 identifies as "the same vehicle at seeds 0 and 4" — one attacker vehicle, one attack family, under 11 placements whose variation Section 4.1 says "does not represent independent traffic samples." The abstract and conclusion nonetheless report that "a reset lost 9.1 per cent of attacker revocations" unqualified. Either report the completed campaign across the four families and both traffic periods, with eligible counts and intervals per family, or restate the headline claims as a demonstration of exactness and cost rather than of effect size.
-
The scheme's own privacy cost is never measured, and the prior work that avoids it is not engaged (Section 1; A2, A4, A5). Privacy motivates the paper, and Section 1 rejects V-TRACE partly because "the authors do not analyse the privacy cost" — but IV-POC is given no such analysis either. The gateway learns the link (A2), the packet carries signatures under both pseudonym keys so that anyone reading the infrastructure link can link them (A4), and A5 requires that "a deployment must refuse a new pseudonym until a transfer receipt exists," which turns every pseudonym change of every vehicle into a linking event at a trusted third party; 396 of the 454 transfers carried no evidence at all. None of this is quantified with the anonymity-set, entropy or traceability measures Section 2.2 attributes to this literature [4]–[6], [20]. Miranda and Rodrigues achieve a transfer in which no party learns the link, so IV-POC's privacy should be stated relative to that, and CoPRA cited as prior art.
Minor Comments
- Tables and figures are numbered out of sequence (Table 5 in Section 4.3 precedes Table 4 in 4.4 and Table 3 in 4.5; Figs 5, 4, 3 and 6 likewise), and Figs 3 to 6 carry captions but no figures.
- Section 4.2 enumerates seven opening attempts per case, but 5,720 checks over 520 cases implies eleven; state what the eleven are.
- The abstract's claim that continuity "halved the honest false-flag rate under a fixed threshold" holds only under the zero-evidence control (Section 4.3), and Table 4 shows 0 of 99 honest vehicles falsely revoked in every arm.
- Section 4.3 attributes the reset's false-flag penalty to lost inter-message history, so a post-change grace period on the sudden-appearance check — as Section 3.4 already grants sequence and nonce history — is the cheaper baseline and should be tested.
- The detector constants of Section 3.2 (decay 0.15, 0.03, 0.02 per second; revocation at 1.2, 1.6, 1.8; witness vote 2.0) are unjustified and never varied, although the denominator of eq. (2) is a direct function of them.
- All 11 placements assess the same 401 observations and learn the same 0.65 threshold from 239 scores (Sections 4.1, 4.3), so explain how placement decides the two reset failures of Section 4.4.
- Table 4 reports only the witnesses-delivered condition; give the suppressed revocation counts there too, since Section 4.4 concludes the benefit depends on corroboration.
- The fit of 423 bytes per observation (Fig. 6) comes from transfers carrying 2 to 12 observations, so it needs an interval and should not be extrapolated to long-lived records.
- Since A5 makes a transfer a precondition of every pseudonym change, estimate the aggregate load on the authorities at realistic rotation rates, not only the per-transfer cost of Table 3.
- Clarify whether the 1,471-byte frozen state lies inside the 8,360-byte capsule (Section 4.5), as the components do not reconcile with the 9,231-byte median packet.
- Reference [25] and Section 2.4 name the scheme RPRRep; the cited paper names it RPRep.
- The comparison with V-TRACE [8] on the same traces, deferred to Section 5, is the experiment that would substantiate preferring transfer over inference.