{"id":"07cadc4b-04eb-4389-89a2-6af8942e5b4c","arxiv_id":"1908.04998","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"IESNN combines probabilistic ranking, adversarial learning, and neural weight updates with searchable encryption, claiming near O(log N) query complexity and low overhead, but without a complete algorithm or implementation.","lead":"This paper sketches an encrypted cloud search system that uses neural networks to rank documents and update search indexes without decrypting data. It claims faster queries and automatic updates, but the algorithms are only described in outline and no working code or proof is provided.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(log N) query claim rests on an unproven and mathematically questionable equivalence between random-query ranking and actual-query ranking; the paper's own Fig. 3 caption concedes that precision is unstable.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: §3.1 asserts without proof that the probabilistic ranking is close to the actual-query ranking. My reading confirms this is the point on which the O(log N) query complexity claim depends, and it is the least secure part of the paper. The paper does present a concrete ASPE-style encryption construction in §2.2 and Eq. (1) correctly computes ciphertext inner products, which is real credit; it also describes an architectural loop of adversarial learning and weight updates. However, none of that supplies the missing mathematical link between random-query ordering and user-query ordering, nor does it provide an algorithm or experiment that demonstrates logarithmic search with acceptable precision. The paper's own Fig. 3 caption acknowledges non-strict ordering and query-dependent precision, which is direct evidence that the central assumption fails under the paper's own conditions. Therefore I see no reason to change the reader's rejection: the direction may be worth exploring, but the central claim is unsupported as written. I would keep the verdict at REJECT, i.e., no adjustment to the reader's verdict.","tokens_in":7192,"tokens_out":4499,"duration_ms":52177,"concrete_test":"Placeholder","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in §3.1 is that after performing 'a sufficient amount of random queries' and sorting index vectors by Σ_j I_iᵀQ_j, 'the index ranking obtained by probabilistic learning is close to the ranking in the actual query.' Algebraically, Σ_j I_iᵀQ_j = I_iᵀ(Σ_j Q_j), so the whole ordering is determined by a single direction R = Σ_j Q_j in the feature space. Nothing in the paper shows that ranking by this fixed random direction approximates ranking under an arbitrary user query q; for any fixed R one can construct queries for which the two rankings disagree substantially. The asserted closeness is stated without proof, bound, or error analysis, and it is exactly the step needed for the claimed O(log N) search: no algorithm is provided that maps a user query to a logarithmic-length path in this ordering, and no complexity analysis accounts for the cost of building or maintaining such an ordering. The paper's own Fig. 3 caption concedes that precision 'is high or low' depending on whether the query is 'popular' and that 'the probabilistic ranking of the index vector is not strictly ordered' — i.e., the assumed closeness fails in the reported experiments. Since the abstract's headline improvements (near-logarithmic query complexity and better-than-previous performance) all depend on this unproven step, the central argument lacks support. The adversarial and weight-update modules in §3.2–3.3 are described only at a high level, with generic equations and no convergence or precision guarantee, so they do not repair the gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IESNN, a framework that combines probabilistic learning, adversarial learning, and automatic weight update with searchable encryption. The claimed contributions are a probabilistic ranking of encrypted index vectors that brings query complexity closer to O(log N), cloud-side automatic weight updates without expensive communication, and low computation/communication overhead compared with prior searchable encryption schemes. The manuscript describes a system architecture, a sorting network based on random queries, a searching adversarial network (SAN), a weight update network (WUN), and an experimental comparison on a collection of 400 IEEE INFOCOM papers.","tokens_in":7528,"tokens_out":2848,"duration_ms":31535,"significance":"If the central claims were established, the paper would address a real problem: reducing the computation and communication cost of dynamic searchable encryption while preserving privacy. The paper also usefully identifies the opportunity to combine machine learning ideas with encrypted search and gives a concrete ASPE-style encryption construction in Section 2.2. However, the load-bearing assertions, especially the probabilistic-ranking closeness assumption and the O(log N) query complexity claim, are stated without proof or algorithmic specification, and the experimental evidence is too thin to validate them. The manuscript is better viewed as a high-level proposal than as a verified scheme; in its current form it does not provide the formal support expected for the claimed results.","major_comments":[{"comment":"The central ranking assumption is unproved and is not supported by the paper's own experiments. The text states that after sorting index vectors by sum of inner products with random queries, 'the index ranking obtained by probabilistic learning is close to the ranking in the actual query.' Algebraically, sum_j I_i^T Q_j = I_i^T (sum_j Q_j), so the entire ordering is determined by one fixed direction R = sum_j Q_j in the feature space. Nothing in the paper shows that ranking by this fixed random direction approximates ranking under an arbitrary user query q; for any fixed direction one can construct queries for which the two rankings differ substantially. The caption of Fig. 3 itself concedes that 'the probabilistic ranking of the index vector is not strictly ordered' and that precision is 'high or low' depending on how popular the query is. Since the abstract's near-logarithmic complexity claim depends on this assumption, the central argument lacks the necessary support.","section":"§3.1, Eq. (1)"},{"comment":"No algorithm is actually given for the claimed O(log N) probabilistic query. The text says that probabilistic query 'brings the query complexity closer to O(log N)', but it does not specify how a user query is mapped to a logarithmic-length search path in the probabilistic ranking, nor does it provide a complexity analysis accounting for the cost of building and maintaining the sorted order. Figure 3 reports only a qualitative comparison of numbers of retrieved indexes; it does not provide the algorithmic steps or a formal complexity bound that would let a reader verify the O(log N) claim.","section":"§3.1, 'probabilistic query algorithm'"},{"comment":"The automatic weight-update mechanism is underspecified at a load-bearing point. The equations use net_j(t) without defining it, and the update principle introduces Delta I_alpha and Delta I_beta without specifying how the increment is computed. The 'local homomorphism' argument asserts (I_alpha + Delta I_alpha)^T M is approximately I_alpha^T M + Delta I_beta^T M, but no relation between Delta I_alpha and Delta I_beta is given, so the approximation is not derivable from the preceding definitions. Because automatic in-cloud update without communication is one of the paper's headline contributions, this gap is substantial.","section":"§3.3, Eqs. (3) and (4)"},{"comment":"The searching adversarial network is described only at a high level, and Eq. (2) is not connected to the described update procedure. The minimax objective in Eq. (2) is a standard GAN-style formulation, but the text says SAN 'only require simple residual calculations and index sorting floating steps' and gives no derivation of those steps from Eq. (2). There is also no convergence analysis and no argument that the 'optimal game equilibrium' is reached in the encrypted setting. This matters because the claim that cloud-side automatic updates preserve search quality depends on SAN behaving as described.","section":"§3.2, Eq. (2)"},{"comment":"The experimental evaluation is not sufficient to support the claimed improvements. The only experiment uses a single small dataset (400 papers, 2,000 keywords), reports no numerical precision or recall values, gives no error bars or statistical tests, and does not measure communication or computation overhead despite those being central claims in the abstract. The caption mentions averaging 100 experimental results, but without the actual numbers and variance, the reader cannot assess whether the differences in Fig. 3 are meaningful.","section":"§3.1 and Fig. 3"}],"minor_comments":[{"comment":"There are typographical issues, such as 'Techno logy' in the affiliation and 'po ssible' in the abstract; a careful proofread is needed.","section":"Author affiliation and abstract"},{"comment":"The variables x and y in Eq. (2) and the distributions pi(x) and pq(y) are not formally defined before the equation, which makes the objective hard to interpret.","section":"§3.2, Eq. (2)"},{"comment":"The checkmark criteria such as 'high-precision query' and 'high-quality ranked search' are not quantitatively defined, so the comparison in Table 1 is difficult to verify.","section":"Table 1"},{"comment":"The activation functions sgn and satlins are used without definition or reference, which is a barrier to reproducing the proposed update rules.","section":"§3.3"},{"comment":"The phrase 'query-learning-update-learning-query self-attention loop' is vague and not tied to the self-attention mechanism in the cited reference [15].","section":"§3.4"}],"recommendation":"reject","confidential_remarks":"The paper's central claims, particularly the O(log N) query complexity and the cloud-side automatic update guarantee, are not backed by the necessary formalization, algorithm specification, or experiments. The probabilistic-ranking assumption is stated as a definition-like assertion and is contradicted by the authors' own Fig. 3 caption. The manuscript may serve as a vision or position paper, but as a research contribution in a security journal it requires either a rigorous proof of the ranking closeness property with explicit bounds, or a substantial revision that removes these unsupported claims. Given the scope of the missing technical content, I cannot recommend acceptance or minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen et al. propose IESNN, a scheme that combines searchable encryption with neural-network-style modules: probabilistic ranking of index vectors via sums of inner products with random queries, a searching adversarial network (SAN), and a weight update network (WUN) meant to let the cloud update index weights without talking to the data owner. The combination is genuinely new as far as I know, and the paper deserves credit for asking whether ML can help where tree-based and matrix-based SE schemes are expensive. The dynamic-update ambitions are also reasonable.\n\nThe problem is that the paper does not actually back any of its headline claims. The abstract promises query complexity closer to O(log N) and better performance than prior work. What the reader gets is a high-level description of the modules, a figure, and a caption that admits the precision is unstable. No concrete algorithm for O(log N) search is given, and no complexity analysis accounts for building or maintaining the ordering that the probabilistic ranking is supposed to produce.\n\nThe stress-test note is right, and the issue is worse than just a missing proof. Summing inner products with many random queries yields I_i · (Σ_j Q_j). So the entire ranking is determined by one fixed random direction. Nothing in the paper suggests that direction will match an arbitrary user query. The authors assert closeness in §3.1; their own Fig. 3 caption concedes that ranking is 'not strictly ordered' and that precision varies with query popularity. That is a load-bearing assumption, and it fails in their own reported experiments.\n\nThe adversarial and weight-update modules are sketched in generic equations (2)–(4) that don't specify how the increment is computed or how convergence is guaranteed. There is no code, no data, no numeric results. The paper reads as a research proposal, not as a demonstrated system. If a concrete algorithm and a reproducible experiment were added, the direction might become publishable.\n\nWho is this for? Someone writing a survey of speculative ML-SE hybrids might cite it, and a reading group could have a fun discussion about why the ranking trick fails. But it should not be sent to serious peer review in its current form. I would desk-reject it and invite the authors to come back with a proof of concept.","headline":"An intriguing but unsubstantiated proposal for ML-based searchable encryption; the central O(log N) claim collapses without a ranking guarantee, so it doesn't yet merit peer review.","tokens_in":8055,"tokens_out":2561,"would_cite":false,"duration_ms":26629,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Probabilistic ranking of encrypted indexes, learned from random queries, can bring encrypted search close to O(log N) query time while updating index weights in the cloud.","keywords":["searchable encryption","probabilistic ranking","encrypted search","neural networks","top-k retrieval","dynamic index update","cloud security","adversarial learning"],"falsifier":"Take a document set with known relevance judgments, build the sorting network exactly as described, then issue real user queries of varying popularity. If the precision of top-k retrieval from the probabilistic ranking falls below a linear-scan baseline whenever the query is not popular, or if precision collapses in the k-regions identified in the paper's Fig. 3, the central claim is false.","tokens_in":6939,"feed_emoji":"🔐","tokens_out":6828,"duration_ms":60810,"temperature":0.7,"pith_summary":"This paper tries to establish that machine learning can fix the two enduring bottlenecks of searchable encryption: slow top-k queries and costly dynamic index updates. It proposes interpretable encrypted searchable neural networks (IESNN), in which a sorting network learns a probabilistic ranking of encrypted index vectors from many uniformly random queries, and the cloud then answers top-k queries by looking up that ranking, with query complexity approaching O(log N). The same construction adds adversarial and weight-update networks that adjust index weights inside the cloud, so indexes can track new data without the user downloading, decrypting, and re-uploading them. If the claim holds, encrypted cloud search becomes practical for large, changing datasets, and ciphertext itself can serve as training data. The evidence is a small real-world experiment (400 documents, 2,000 keywords) whose own figure caption concedes that precision is not stable across all values of k.","feed_headline":"Encrypted search can approach O(log N) with probabilistic ranking","feed_subtitle":"Random-query sorting and cloud-side weight updates aim to cut the cost of dynamic encrypted search.","key_machinery":"The load-bearing object is the sorting network built from encrypted index vectors. Each document index is a weighted vector over the keyword dictionary, padded with pseudo-keywords and encrypted by splitting the vector and multiplying with random invertible matrices so that inner products can still be computed on ciphertext. During setup the network issues many random queries $Q_j \\sim U(-\\sigma\\sqrt{3}, \\sigma\\sqrt{3})$, computes the sum $\\sum_j I_i^T Q_j$ for each index, and sorts indices by this score; that precomputed order is the probabilistic ranking used for queries. Updates are carried by two auxiliary networks: a searching adversarial network computes residual weight changes and ranking shifts from top-k feedback, and a weight-update network applies those changes through vector and matrix operations that are locally homomorphic to plaintext increments.","core_discovery":"The paper's central claim is that ranking encrypted index vectors by the sum of their inner products with many random query vectors gives an ordering close to the ordering a real user query would produce, and that this probabilistic ranking supports top-k retrieval in time close to O(log N). On top of this, a searching adversarial network and a weight-update network (combining backpropagation with a discrete recurrent network) let the cloud update index weights automatically, so dynamic operations add, delete, change, and search for keywords without expensive communication. The authors state that this combination outperforms tree-based search [3,13] and matrix-based multi-keyword ranked search [1,6], and that the update operations are performed in the cloud with only small amounts of data touched.","pith_inferences":["If the probabilistic-ranking assumption is formalized, the natural next step is an error bound: for a given number of random queries and query distribution, how close the learned ranking is to the optimal ranking can likely be quantified with concentration inequalities; the paper does not attempt this.","The sorting network's reliance only on inner products suggests the idea could transfer to any encrypted inner-product scheme, not just the particular matrix-splitting encryption used here.","A practical deployment would need a privacy analysis of the update residuals: returning top-k feedback to the adversarial network reveals ranking information that a semi-trusted cloud could exploit; the paper asserts security in the known-background model but does not analyze this leakage.","The approach could combine with tree indices: probabilistic ranking could choose which subtree to descend first, potentially making tree search robust in multi-owner settings where ordinary balanced trees degrade."],"forward_implications":["Top-k retrieval no longer needs a full scan or a balanced tree: the precomputed probabilistic ranking lets the cloud read off the top-k indexes, bringing query complexity toward O(log N).","Dynamic updates (adding or removing documents or keywords) can be executed in the cloud by adjusting index weights, eliminating the download-decrypt-update-upload cycle of earlier schemes.","The index itself becomes a training signal: ciphertext-based sorting can produce rankings consistent with plaintext analysis, so encrypted data can feed machine-learning pipelines.","Precision depends on how 'popular' the query vector is; the authors report that unpopular queries lower precision because the probabilistic ranking is not strictly ordered."],"supporting_citations":[{"why":"Defines the MRSE baseline the paper compares against and the known-background threat model.","marker":"[1]"},{"why":"Supplies the adversarial minimax objective that the searching adversarial network adapts.","marker":"[2]"},{"why":"Provides the multi-data-owner ranked-search baseline and one of the tree-based methods used in performance comparison.","marker":"[3]"},{"why":"Supplies the backpropagation network component of the weight-update network.","marker":"[4]"},{"why":"Provides the dynamic-dictionary extension and matrix-based baseline (MKQE) used in comparisons.","marker":"[6]"},{"why":"Supplies the discrete neural-network update rule used in the weight-update network.","marker":"[7]"},{"why":"Provides the asymmetric scalar-product-preserving encryption used to compute inner products on ciphertext.","marker":"[12]"},{"why":"Provides the secure dynamic tree-based search baseline whose query cost IESNN aims to beat.","marker":"[13]"}],"fun_headline_variants":["Probabilistic ranking speeds encrypted search to near O(log N)","Cloud-side neural nets auto-update encrypted search indexes","Random query vectors rank encrypted data with near-log time","Encrypted search gets cloud-side weight updates, hits O(log N)","Neural nets make encrypted search dynamic and nearly log-time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole design rests on the assumption that summing inner products with many uniformly random query vectors produces an index ranking close to the ranking a real user query would produce; the paper states this closeness without proof.","fun_headline_variants_meta":{"raw":{"variants":["Probabilistic ranking speeds encrypted search to near O(log N)","Cloud-side neural nets auto-update encrypted search indexes","Random query vectors rank encrypted data with near-log time","Encrypted search gets cloud-side weight updates, hits O(log N)","Neural nets make encrypted search dynamic and nearly log-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000759,"raw_usage":{"total_tokens":3317,"prompt_tokens":836,"completion_tokens":2481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":2399}},"tokens_in":452,"tokens_out":2481,"duration_ms":20555,"temperature":1.0,"reasoning_tokens":2399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:26:12.888659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a document set with known relevance judgments, build the sorting network exactly as described, then issue real user queries of varying popularity. If the precision of top-k retrieval from the probabilistic ranking falls below a linear-scan baseline whenever the query is not popular, or if precision collapses in the k-regions identified in the paper's Fig. 3, the central claim is false.","supporting_citations":[{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Defines the MRSE baseline the paper compares against and the known-background threat model."},{"cited_title":"Journal of Systems and Soft- ware 137(3), 380–395 (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the multi-data-owner ranked-search baseline and one of the tree-based methods used in performance comparison."},{"cited_title":"C ognitive Science 30(4), 725–731 (2006)","cited_arxiv_id":null,"evidence_quote":"Supplies the backpropagation network component of the weight-update network."},{"cited_title":"Future Generation C omp","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic-dictionary extension and matrix-based baseline (MKQE) used in comparisons."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete neural-network update rule used in the weight-update network."},{"cited_title":"In: ACM SIGMOD 2009","cited_arxiv_id":null,"evidence_quote":"Provides the asymmetric scalar-product-preserving encryption used to compute inner products on ciphertext."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Provides the secure dynamic tree-based search baseline whose query cost IESNN aims to beat."}],"review_version":1}