Bot Adversarial Attack - lkb1324/Bot-Adversarial-Attack-and-Detection GitHub Wiki
Bot检测工作
[TDSC’21] He X, Gong Q, Chen Y, et al. DatingSec: Detecting malicious accounts in dating apps using a content-based attention network.
一种混合模型用于检测dating app中的恶意用户。实验设计部分针对模型鲁棒性进行了对对抗样本的检测。通过增加不同比例的manipulated sample分析模型效果的下降情况。注意区分malicious user和poisoned user: poisoned user:注入误导信息但是被标记为合法用户,用以扰乱检测训练 malicious user:正常的恶意用户
Bot检测中基于community detection的工作
[WWW'23] Li S, Zhao C, Li Q, et al. BotFinder: a novel framework for social bots detection in online social networks based on graph embedding and community detection[J].
采用基于社区检测的无监督方法对检测结果进行修正,进一步提升模型性能。
[TDSC'24] Liu F, Yang C, Li Z, et al. Accou2vec: A Social Bot Detection Model Based on Community Walk[J].
对社交图进行社区检测,将其划分为多个子图,以此辨别并切断机器人与人类节点之间的联系
[Inf Sys'22] Pham P, Nguyen L T T, Vo B, et al. Bot2Vec: A general approach of intra-community oriented representation learning for bot detection in different types of social networks[J].
对Node2Vec的采样策略进行改进,通过将节点邻域的采样限制在节点所处的社区内有效保留了节点的社区内部结构属性
[TOIS'20]Mendoza M, Tesconi M, Cresci S. Bots in social and interaction networks: detection and impact estimation[J].
不是标准的community detection但是利用了pairwise similarity和基于此的proximity graph进行label propagation。
Bot检测工作脆弱性分析
[UBMK'20] Are we secure from bots? Investigating vulnerabilities of Botometer.
bot数据集构建与算法vulnerability检测:利用研发的bot数据集而非公开数据集验证流行的基于机器学习的bot分类器Botometer的安全性。构建的机器人确实能够绕过botometer的检测,但是只揭露了现有detection的短板,没有给出进一步的分析与建议
Bot对抗攻击
[WebSci’19] Cresci S, Petrocchi M, Spognardi A, et al. Better safe than sorry: an adversarial approach to improve social bot detection.
设计算法模仿bot进化,并根据进化后的bot特征提出检测方法,目的是通过预测bot的进化模式,主动提升检测算法能力
[MM'23] Wang L, Qiao X, Xie Y, et al. My Brother Helps Me: Node Injection Based Adversarial Attack on Social Bot Detection.
基于单点注入的黑盒攻击,能够有效攻击一类利用节点属性与图结构进行bot检测的GCN模型,通过合理设计替代模型,并利用对抗攻击的可转移性实现黑盒攻击,单点注入保证了攻击的不可感知
[Technologies2020] Kantartopoulos P, Pitropakis N, Mylonas A, et al. Exploring adversarial attacks and defences for fake twitter account detection[J].
针对machine learning classifier的bot detection算法的adversarial attack的分析与对策 提出两种针对分类器的对抗攻击,并提出解决对策:检测数据集中存在的有毒数据
[KDD Explorations'21] Liu N, Du M, Guo R, et al. Adversarial attacks and defenses: An interpretation perspective[J]. ACM SIGKDD Explorations Newsletter'21.
从机器学习与深度学习的可解释性角度理解对抗性攻击与防御。本文将深度学习模型的可解释性分为两个层次:特征级解释与模型模块功能级解释,将特征级与模块功能级的的模型原理转化为人类能够理解的术语,同时对对抗性攻击进行分类分析,从模型解释的角度理解对抗性攻击与防御的最新发展
良性Bot应用
[CHI'20] Taewan Kim, Mintra Ruensuk, and Hwajung Hong. 2020. In Helping a Vulnerable Bot, You Help Yourself: Designing a Social Bot as a Care-Receiver to Promote Mental Health and Reduce Stigma. Proc. of ACM CHI, 2020.
一个基于Github数据集与XGBoost的开源项目GFI推荐机器人,利用github API完成数据集定期更新维护和模型迭代训练,提供推荐内容调整的超参数并完善用户应用程序。 takeaways:是一个基于平台功能,用户属性与问题需求的良性机器人,其功能与平台需求紧耦合。
[ESEC/FSE'22] Hao He, Haonan Su, Wenxin Xiao, Runzhi He, and Minghui Zhou. GFI-Bot: Automated Good First Issue Recommendation on GitHub. Proc. of ESEC/FSE, 2022.
利用社交机器人模拟抑郁症患者,弥补与抑郁症患者本人直接接触的social contact形式对抑郁症干预治疗的效果研究,并规避真实接触中可能存在的风险用于提供社会接触的治疗支持