{"id":"b720a7bd-162e-4ac4-9fce-fbe43074984a","arxiv_id":"1908.03446","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper demonstrates a distributed simulated annealing algorithm over a blockchain for estimating a binary logit mode choice model, reporting parameters consistent with a centralized version, but with acknowledged privacy limitations.","lead":"A distributed choice modeling tool is presented that keeps personal data on participants' devices and shares only computed model evaluations over a blockchain. It is a small-scale demonstration of using blockchain and distributed optimization for travel behavior modeling.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Honest-but-curious chief can infer workers' observations from scalar log-likelihoods; the paper's own conclusion concedes this, so the privacy-aware claim is unsupported.","rationale":"The reader's weakest assumption already identified the same load-bearing concern: the honest-but-curious chief can infer worker observations from the decrypted local log-likelihood values. The distributed objective is exactly the sum of local log-likelihoods, so the numerical demonstration of consistency with the centralized objective is sound in a narrow sense; however, the advertised privacy property fails under a standard adversary model, and the authors' own conclusion concedes this. A secondary correctness concern is that Algorithm 1 draws the perturbation epsilon once outside the sampling loop (line 9), which would make the published pseudo-code unable to explore the parameter space as an annealing algorithm; if the GitHub implementation matches the pseudo-code, the reported estimates are not a valid maximum-likelihood result. Since the reader already conditioned acceptance on differential privacy and algorithm corrections, this stress-test does not change the conditional verdict; it strengthens the rationale for requiring those conditions.","tokens_in":7499,"tokens_out":10085,"duration_ms":118007,"concrete_test":"Simulate an honest-but-curious chief against one worker, using either the actual 61-observation subset from the SP survey or a synthetic copy. Have the chief query l_w(beta) for 10^4 to 10^5 beta vectors spanning the plausible parameter range, then reconstruct the worker's triples by minimizing the squared difference between observed and predicted l_w over free parameters {Delta c_n, Delta t_n, y_n}_{n=1}^{61}, initialized from multiple random starting points and using the closed-form derivatives. If the recovered attributes and choices match the true values to within the resolution of the SP survey (loss reaches the numerical noise floor), the privacy attack is confirmed. A cheaper analytical check is to use the derivative identity dl_w/d beta_c = sum_n (y_n - sigma(u_n)) Delta c_n and verify that successive derivatives reveal all moments of the empirical distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central privacy guarantee is undermined by the chief's ability to query worker log-likelihoods. Each worker sends l_w(beta) = sum_n [y_n u_n - log(1 + exp(u_n))], where u_n = beta_a + beta_c * Delta c_n + beta_t * Delta t_n and (Delta c_n, Delta t_n, y_n) are the worker's personal observations. Because the chief controls the model updates, it can obtain l_w for many beta vectors. The derivatives of l_w with respect to beta are moments of the empirical distribution of these triples, so in the limit of enough queries the chief can recover each worker's (Delta c, Delta t, y) to arbitrary precision, not merely approximately. The authors acknowledge this in the Conclusions: 'Chief nodes knows the model and have access to the decrypted worker evaluations, hence if the chiefs are honest but curious (in the best case) they may be able to infer personal observations of workers by using the evaluations.' Since the chief is the most powerful node, this admission means that the architecture prevents direct sharing of raw data but does not provide privacy against the standard honest-but-curious adversary. Therefore the title claim 'privacy-aware' and the abstract's statement that transactions are 'secure and private' are not supported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed framework for estimating binary logit mode choice models over the BSMD blockchain, using a distributed version of simulated annealing (dSA). Workers compute local log-likelihood contributions on their private observations and send only scalar evaluation values to a chief node; a case study with 246 stated-preference observations, one chief, and four Raspberry Pi worker nodes reports parameter estimates and message latencies. The authors claim that the estimated model parameters are consistent and reproducible, and that the architecture keeps personal raw data private and transactions secure.","tokens_in":7753,"tokens_out":3371,"duration_ms":33776,"significance":"If the claims held, the paper would provide a useful template for privacy-preserving distributed estimation of discrete choice models, with a reproducible implementation on real hardware. The authors are transparent about several limitations, which is commendable. However, as presented, the central privacy claim is contradicted by the authors' own admission in the Conclusions, and the simulated annealing implementation contains a proposal-generation flaw that undermines the reported convergence result. The consistency claim is also supported only by an internal comparison with a centralized version of the same algorithm, not by an independent estimator. These issues are load-bearing for the paper's main claims, so substantial revision is needed.","major_comments":[{"comment":"The proposal step is degenerate: epsilon is sampled once before the temperature loop (line 9) and then reused for every proposal (line 13), so beta_new is always beta plus the same constant vector. This is not a simulated annealing random walk with local proposals; the search reduces to a deterministic shift with random acceptance, and the reported convergence behavior does not demonstrate a working distributed SA. The epsilon must be resampled inside the inner loop.","section":"Section 4, Algorithm 1, lines 9 and 13"},{"comment":"The Conclusions state: 'Chief nodes knows the model and have access to the decrypted worker evaluations, hence if the chiefs are honest but curious (in the best case) they may be able to infer personal observations of workers by using the evaluations.' This admission directly contradicts the abstract's claim that transactions are 'secure and private' and the title's 'Privacy-Aware'. Because the chief controls the sequence of beta vectors and receives l_w(beta) for each, it can differentiate l_w with respect to beta and recover moments of each worker's empirical distribution of (Delta c, Delta t, y); in the limit of many queries, exact recovery of the underlying observations is possible. The privacy guarantee is therefore unsupported for the standard honest-but-curious adversary, and the architecture at best prevents direct raw-data sharing. The authors should either add a formal privacy mechanism (e.g., differential privacy) or explicitly re-scope the privacy claim to exclude honest-but-curious chiefs.","section":"Section 6, Conclusions"},{"comment":"The verification consists solely of comparing the distributed algorithm with a non-distributed version of the same simulated annealing algorithm. This is an internal consistency check, not a validation against an independent estimator such as standard maximum likelihood software, and no multiple-run variability, random-seed analysis, or convergence diagnostic is reported. Consequently, the abstract's claim that the estimated parameters are 'consistent and reproducible' is not established. The authors should provide an external benchmark and repeated runs with different random seeds.","section":"Section 4.2 and Table 1"}],"minor_comments":[{"comment":"The text says 'each RPI runs a script similar to Algorithm 1', but Algorithm 1 is the chief's algorithm; workers should run a script similar to Algorithm 2. Please correct this mismatch, which affects reproducibility.","section":"Section 4.1, System setup"},{"comment":"Please fix typographical errors: 'state preference' should be 'stated preference'; 'likehood' should be 'likelihood'; 'Canda' should be 'Canada'; 'massage' should be 'message'; and 'conversion' in Section 4.2 should be 'convergence'.","section":"Abstract and Section 3"},{"comment":"The Introduction states that 'all the information transactions are recorded on the blockchain', but Section 3 says that evaluation values of the objective function are not published in the ledger to avoid reverse-engineering. Please clarify which transactions are recorded and why the two statements are not contradictory.","section":"Section 3, Distributed process"},{"comment":"The standard errors in Table 1 are presented without any explanation of how they were computed from the simulated annealing output. Please state the estimation procedure for these standard errors.","section":"Table 1"},{"comment":"The statement that an attacker 'could do is to only approximate the personal observations' is in tension with the stronger admission in the Conclusions that honest-but-curious chiefs 'may be able to infer personal observations of workers'. Please reconcile these two statements.","section":"Section 5, Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper is based heavily on the authors' own prior BSMD framework, and the new contribution here is the distributed simulated annealing and the case study. The proposal-generation flaw and the privacy contradiction are both fixable in principle—resampling epsilon and re-scoping the privacy claims would address the technical core. However, the authors should be asked to either provide a meaningful external validation or substantially weaken the consistency and privacy claims. I see no indication of misconduct; the main issue is overclaiming relative to what is demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on 1908.03446.\n\nWhat's new: they take their own BSMD blockchain and show that a binary logit mode choice model can be estimated in a distributed way, with workers sending only scalar log-likelihoods and their raw data never leaving their devices. The case study runs a distributed simulated annealing over one chief and four Raspberry Pi workers on 246 observations and gets the same parameters as the centralized version. That reproducibility check is the paper's real content: it shows the plumbing works as a template. The github repo and setup instructions are a plus.\n\nWhat's soft: the headline privacy claim does not survive contact with their own conclusion. They admit that an honest-but-curious chief can infer individual observations from the decrypted worker evaluations. The stress-test note is right, and it's worse than approximate: the chief controls the beta proposals, so it can query each worker's l_w at many points. The derivatives of l_w with respect to beta are empirical moments of that worker's (Delta c, Delta t, y) distribution, so a few hundred queries recover the worker's data to machine precision. So the architecture prevents casual data sharing but does not provide privacy against the standard adversary named in the paper. Also Algorithm 1 samples epsilon once, outside the annealing loop, so the proposal is a fixed step; the random walk is actually a deterministic drift. That's a bug, likely a typo, but it means the distributed algorithm is not doing what the text says. The consistency check compares dSA to the same centralized SA, not to a conventional MLE; the standard errors are reported without explanation of how they were computed. These do not sink the demonstration, but they cap its value.\n\nProportion: this is a proof-of-concept, and as such it's honest and useful. The limitation paragraph in the conclusions is a credit; they flag the chief-inference problem themselves. The paper doesn't overclaim convergence theory; it claims consistency in the narrow sense of matching the centralized run, and that holds.\n\nWho should read it: people working on privacy-preserving travel data collection, and anyone curious about blockchain as a coordination layer for federated estimation. It's not a strong privacy contribution, but it's a concrete example worth knowing.\n\nRecommendation: send it to peer review. A serious reviewer can push them to fix the epsilon bug, add differential privacy or clearly reframe the privacy claim, and validate against a standard MLE. The idea deserves a round of revisions, not a desk reject.","headline":"A useful proof-of-concept for distributed choice model estimation, but the privacy claim is undercut by the authors' own honest-but-curious chief admission and the dSA code has a deterministic-step bug.","tokens_in":8237,"tokens_out":2118,"would_cite":false,"duration_ms":20757,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Distributed blockchain framework estimates mode-choice models while personal raw data stays on devices.","keywords":["distributed choice modelling","blockchain","simulated annealing","mode choice","privacy","maximum likelihood","smart mobility data-market"],"falsifier":"Give a chief node the decrypted local log-likelihoods from a single worker over the full sequence of parameter proposals, together with the model and the worker's data size, and test whether a maximum-likelihood or equation-solving attack recovers the worker's individual observations; recovery of even one observation falsifies the privacy claim.","tokens_in":7292,"feed_emoji":"🚆","tokens_out":4904,"duration_ms":49003,"temperature":0.7,"pith_summary":"This paper tries to establish that discrete choice models, specifically a binary logit mode choice model, can be estimated in a distributed, privacy-preserving way when personal raw data never leaves participants' devices. It presents a distributed version of simulated annealing running over the Blockchain-based Smart Mobility Data-market (BSMD), where workers send only their local log-likelihood values to a chief node. The paper claims the estimated parameters match a centralized version of the same algorithm, making the procedure consistent and reproducible. If true, this provides a template for privacy-aware mobility modelling that avoids central data collection. However, the paper itself concedes that a chief node seeing decrypted evaluations might infer personal observations, which undermines the full privacy claim.","feed_headline":"Mode-choice models estimated while raw data never leaves phones","feed_subtitle":"Distributed simulated annealing matches central estimation using only local likelihood values from workers.","key_machinery":"The mechanism that carries the argument is the separation between local evaluation and global aggregation. Each worker $w$ holds $N$ personal observations and computes its local log-likelihood $l_w(\\vec{\\beta}) = \\sum_{n=1}^{N}\\ln(y_{a,n}P(a_n;\\vec{\\beta}) + y_{tr,n}P(tr_n;\\vec{\\beta}))$ for the parameter vector sent by the chief; only this scalar travels over an encrypted peer-to-peer channel, while the blockchain records who sent what to whom, not the evaluation value. The chief sums the $l_w$'s into a global log-likelihood $l$ and applies the simulated annealing acceptance rule. This lets any optimisation algorithm that only needs the objective function value be distributed without moving raw data, provided workers can evaluate the function locally.","core_discovery":"The central claim is that a binary logit mode choice model can be estimated in a distributed setting where each worker owns a subset of the observations and never shares them. The chief distributes candidate parameter vectors $\\vec{\\beta}$; each worker evaluates the log-likelihood contribution $l_w$ of its own observations and sends back only this scalar; the chief sums the values and runs a simulated annealing acceptance rule. The reported parameters ($\\beta_a=0.3444$, $\\beta_c=-0.0062$, $\\beta_t=-0.0008$) match those obtained by running the same annealing procedure centrally, so the estimation is described as consistent and reproducible. The framework is positioned as a template for privacy-aware mobility modelling because the raw data stays on the participants' devices.","pith_inferences":["A natural extension is to apply the same local-evaluation/global-aggregation split to models whose log-likelihoods are additively separable over rows, such as survival models or generalized linear models; the architecture does not depend on the transport context.","The honest-but-curious chief attack is the decisive test for the privacy claim: if a chief can reconstruct any worker's observation from the sequence of decrypted likelihoods, the framework reduces to a data-market that still leaks through the objective function.","Because the dSA convergence rests on the same acceptance rule as centralized simulated annealing, the reproducibility claim would be strengthened by running multiple restarts from different seeds and reporting the distribution of final parameters rather than a single run.","The communication cost (about 80 MB per worker) comes from sending one parameter vector per annealing iteration; evaluating a batch of parameter candidates in one message would cut both latency and bandwidth, at the price of a slightly different annealing schedule."],"forward_implications":["If the framework is correct, mode choice models and other maximum-likelihood choice models can be estimated over data held by many private individuals without central data collection or central data storage.","Any estimation algorithm that only consumes objective-function values—not gradients or individual records—can be plugged into the same distributed pipeline, including nested or mixed logit specifications.","The measured latencies (under 0.05 s per message at workers, about 0.15 s at the chief) suggest the approach can run on ordinary mobile equipment, although the roughly 80 MB of messages per worker over 200,000 rounds makes unmetered networks a practical condition.","The blockchain ledger gives participants an auditable record of with whom and when information was shared, which supports accountability in case a node misuses data.","The paper's own conclusion implies that any real deployment must add noise or secure aggregation before the system is safe against a curious chief."],"supporting_citations":[{"why":"Defines the BSMD blockchain framework used as the platform for the distributed estimation.","marker":"[8]"},{"why":"Extends BSMD into the multi-layered framework that supports identity keys and encrypted peer-to-peer distribution.","marker":"[9]"},{"why":"Supplies the stated-preference survey data (246 observations) on which the binary logit model is estimated.","marker":"[21]"},{"why":"Provides the simulated annealing method that the distributed algorithm adapts for maximum-likelihood estimation.","marker":"[22]"}],"fun_headline_variants":["Distributed mode-choice estimation with zero raw data sharing","No raw data off-device: distributed mode choice matches central","Distributed simulated annealing keeps mobility data on-device","Blockchain + local likelihood = private mobility choice","Zero raw data shared: distributed mode-choice matches central"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing privacy premise is that sending a worker's local log-likelihood value to the chief does not let the chief infer that worker's raw observations; the paper itself states this is doubtful for honest-but-curious chiefs.","fun_headline_variants_meta":{"raw":{"variants":["Distributed mode-choice estimation with zero raw data sharing","No raw data off-device: distributed mode choice matches central","Distributed simulated annealing keeps mobility data on-device","Blockchain + local likelihood = private mobility choice","Zero raw data shared: distributed mode-choice matches central"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3183,"prompt_tokens":780,"completion_tokens":2403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":396,"completion_tokens_details":{"reasoning_tokens":2326}},"tokens_in":396,"tokens_out":2403,"duration_ms":18332,"temperature":1.0,"reasoning_tokens":2326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:12:22.629305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give a chief node the decrypted local log-likelihoods from a single worker over the full sequence of parameter proposals, together with the model and the worker's data size, and test whether a maximum-likelihood or equation-solving attack recovers the worker's individual observations; recovery of even one observation falsifies the privacy claim.","supporting_citations":[{"cited_title":"A blockchain framework for smart mobility","cited_arxiv_id":null,"evidence_quote":"Defines the BSMD blockchain framework used as the platform for the distributed estimation."},{"cited_title":"Innovative intercity transport mode: Application of choice preference integrated with attributes nonattendance and value learning","cited_arxiv_id":null,"evidence_quote":"Supplies the stated-preference survey data (246 observations) on which the binary logit model is estimated."},{"cited_title":"A course in simulation","cited_arxiv_id":null,"evidence_quote":"Provides the simulated annealing method that the distributed algorithm adapts for maximum-likelihood estimation."}],"review_version":1}