On Manipulating Weight Predictions in Signed Weighted Networks
Authors: Tomasz Lizurej, Tomasz Michalak, Stefan Dziembowski
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experimental analysis, we first evaluate two benchmarks: (a) the strength of the aforementioned direct attack, and (b) the strength of an indirect attack based on a simple greedy approach. The latter one turns out to be very ineffective. Next, we analyse an improved greedy approach by attacking at a larger scale in every step. This approach, although costly, proves to be often effective. Simulations We conduct a series of simulations on the Bitcoin OTC, Bitcoin Alpha, and RFA Net datasets studied by Kumar et al. (2016). |
| Researcher Affiliation | Academia | Tomasz Lizurej1,2, Tomasz P. Michalak1,2, Stefan Dziembowski1,2 1 University of Warsaw 2 IDEAS NCBR tomasz.lizurej@crypto.edu.pl, tpm@mimuw.edu.pl, stefan.dziembowski@crypto.edu.pl |
| Pseudocode | Yes | Algorithm 1: Direct attack Data: A, T = {t}, G for a A do add an edge (a, t), ω(a, t) = 1 to the graph G end Algorithm 2: Indirect attack Data: A, T = {t}, G sort nodes in A by their fairness score for a sorted(A) do N1 Pred(t) find a neighbor n2 Succ(n1) \ {t} of a neighbor n1 N1 that minimizes the goodness value of t, when adding an edge (a, n2) with weight ω(a, n2) = 1 or ω(a, n2) = 1 add the edge to the graph G Algorithm 3: Modified indirect attack Data: A, T = {t}, G sort nodes in A by their fairness score while i < len(sorted(A)) do a sorted(A)[i] N1 Pred(t) find a neighbor n2 Succ(n1) \ {t} of a neighbor n1 N1 that minimizes the goodness value of t, when adding an edge (a, n2) with weight ω(a, n2) = 1 or ω(a, n2) = 1 edges len min(SCALE len(indegree(n2), MAX, len(A) i) add edges len edges to the graph G i i + edges len end |
| Open Source Code | No | The paper does not provide any concrete access information (e.g., a specific repository link or explicit statement of release) for the source code of the methodology described. |
| Open Datasets | Yes | We conduct a series of simulations on the Bitcoin OTC, Bitcoin Alpha, and RFA Net datasets studied by Kumar et al. (2016). They consist of weighted signed networks with |V | = 3, 700, |E| 24, 000 each, where the proportion of positively weighted edges is 84%. |
| Dataset Splits | No | The paper describes how target nodes were chosen and attack strategies but does not provide specific details on how the datasets were split into training, validation, or test sets in a reproducible manner (e.g., percentages, sample counts, or predefined splits). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments (e.g., exact GPU/CPU models, processor types, or memory amounts). |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., libraries, frameworks, or programming language versions) needed to replicate the experiments. |
| Experiment Setup | Yes | The target, t T, was chosen randomly from those nodes that have relatively high goodness (g(t) 0.50) and a low indegree (0 < indeg(t) < 10). We study two types of the attackers: not-established attackers chosen from relatively newly created nodes with 0 < indeg < 10 and outdeg = 0). This allows for studying Sybil-style attacks; and established attackers chosen from the nodes with outdeg(v) > 5 (and iteratively choosing nodes with fairness f(v) > 0.7). This allows for studying attacks by the nodes whose standing in the network has been built for some time. We simulate three types of attacks: direct attacks a set of attackers A of size k rates directly the target node t T. The pseudocode is presented in Algorithm 1. Each attacker rates t using weight 1; indirect attacks the attackers set A of size k rates the neighbors of the neighbors of the target node, to minimize the goodness part of the FGA of the target node by manipulating fairness of the targets neighbors. The pseudocode of the attack is presented in the Alorithm 2. More precisely, the algorithm implements a greedy approach, where each new edge is used to minimize the goodness of the target node t by minimizing (or maximizing) the fairness of one of the targets neighbors by directly rating the successor of the target s neighbor with an edge of weight 1 or 1. The algorithm performs calculations iteratively on the attackers sorted by the value of their fairness value. mixed attack k1 attacking nodes perform a direct attack and k2 perform an indirect one, where k1 + k2 = k. In Algorithm 3, instead of adding only a single edge in each iteration, as in Algorithm 2, we add a series of new edges. In more detail, we add SCALE = 5 times more Sybil edges than the indegree of the target node (but at most some predefined maximum MAX = 10). |