Spam emails pose a threat to individuals. The proliferation of spam emails daily has rendered traditional machine learning and deep learning methods for screening them ineffective and inefficient. In our research, we ...Spam emails pose a threat to individuals. The proliferation of spam emails daily has rendered traditional machine learning and deep learning methods for screening them ineffective and inefficient. In our research, we employ deep neural networks like RNN, LSTM, and GRU, incorporating attention mechanisms such as Bahdanua, scaled dot product (SDP), and Luong scaled dot product self-attention for spam email filtering. We evaluate our approach on various datasets, including Trec spam, Enron spam emails, SMS spam collections, and the Ling spam dataset, which constitutes a substantial custom dataset. All these datasets are publicly available. For the Enron dataset, we attain an accuracy of 99.97% using LSTM with SDP self-attention. Our custom dataset exhibits the highest accuracy of 99.01% when employing GRU with SDP self-attention. The SMS spam collection dataset yields a peak accuracy of 99.61% with LSTM and SDP attention. Using the GRU (Gated Recurrent Unit) alongside Luong and SDP (Structured Self-Attention) attention mechanisms, the peak accuracy of 99.89% in the Ling spam dataset. For the Trec spam dataset, the most accurate results are achieved using Luong attention LSTM, with an accuracy rate of 99.01%. Our performance analyses consistently indicate that employing the scaled dot product attention mechanism in conjunction with gated recurrent neural networks (GRU) delivers the most effective results. In summary, our research underscores the efficacy of employing advanced deep learning techniques and attention mechanisms for spam email filtering, with remarkable accuracy across multiple datasets. This approach presents a promising solution to the ever-growing problem of spam emails.展开更多
With the rapid growth of the Internet in recent years, the ability to analyze and identify its users has become increasingly important. Authorship analysis provides a means to glean information about the author of a d...With the rapid growth of the Internet in recent years, the ability to analyze and identify its users has become increasingly important. Authorship analysis provides a means to glean information about the author of a document originating from the internet or elsewhere, including but not limited to the author’s gender. There are well-known linguistic differences between the writing of men and women, and these differences can be effectively used to predict the gender of a document’s author. Capitalizing on these linguistic nuances, this study uses a set of stylometric features and a set of word count features to facilitate automatic gender discrimination on emails from the popular Enron email dataset. These features are used in conjunction with the Modified Balanced Winnow Neural Network proposed by Carvalho and Cohen, an improvement on the original Balanced Winnow created by Littlestone. Experiments with the Modified Balanced Winnow show that it is effectively able to discriminate gender using both stylometric and word count features, with the word count features providing superior results.展开更多
In this paper, we propose a new online system that can quickly detect malicious spam emails and adapt to the changes in the email contents and the Uniform Resource Locator (URL) links leading to malicious websites by ...In this paper, we propose a new online system that can quickly detect malicious spam emails and adapt to the changes in the email contents and the Uniform Resource Locator (URL) links leading to malicious websites by updating the system daily. We introduce an autonomous function for a server to generate training examples, in which double-bounce emails are automatically collected and their class labels are given by a crawler-type software to analyze the website maliciousness called SPIKE. In general, since spammers use botnets to spread numerous malicious emails within a short time, such distributed spam emails often have the same or similar contents. Therefore, it is not necessary for all spam emails to be learned. To adapt to new malicious campaigns quickly, only new types of spam emails should be selected for learning and this can be realized by introducing an active learning scheme into a classifier model. For this purpose, we adopt Resource Allocating Network with Locality Sensitive Hashing (RAN-LSH) as a classifier model with a data selection function. In RAN-LSH, the same or similar spam emails that have already been learned are quickly searched for a hash table in Locally Sensitive Hashing (LSH), in which the matched similar emails located in “well-learned” are discarded without being used as training data. To analyze email contents, we adopt the Bag of Words (BoW) approach and generate feature vectors whose attributes are transformed based on the normalized term frequency-inverse document frequency (TF-IDF). We use a data set of double-bounce spam emails collected at National Institute of Information and Communications Technology (NICT) in Japan from March 1st, 2013 until May 10th, 2013 to evaluate the performance of the proposed system. The results confirm that the proposed spam email detection system has capability of detecting with high detection rate.展开更多
近年来,大型语言模型(large language model,LLM)技术兴起,类似ChatGPT这样的AI机器人,虽然其内部设置了大量的安全对抗机制,攻击者依然可以精心设计问答,绕过这些AI机器人的安全机制,在其帮助下自动化生产钓鱼邮件,进行网络攻击.这种...近年来,大型语言模型(large language model,LLM)技术兴起,类似ChatGPT这样的AI机器人,虽然其内部设置了大量的安全对抗机制,攻击者依然可以精心设计问答,绕过这些AI机器人的安全机制,在其帮助下自动化生产钓鱼邮件,进行网络攻击.这种情形下,如何鉴别AI生成的文本也成为一个热门的问题.为了开展LLM生成内容检测实验,从互联网某社交平台和ChatGPT收集了一定数量的问答数据样本,依据AI文本可获得条件的不同,研究提出了一系列检测策略,包含基于在线可获取AI对照样本的文本相似度分析、基于离线条件下使用统计差异性的文本数据挖掘分析、基于无法获得AI样本条件下的LLM生成方式对抗分析以及基于通过微调目标LLM模型本身构建分类器的AI模型分析,计算并比较了每种情况下分析引擎的检测能力.另一方面,从网络攻防的角度,针对检测策略的特点,给出了一些对抗AI文本检测引擎的免杀技巧.展开更多
文摘Spam emails pose a threat to individuals. The proliferation of spam emails daily has rendered traditional machine learning and deep learning methods for screening them ineffective and inefficient. In our research, we employ deep neural networks like RNN, LSTM, and GRU, incorporating attention mechanisms such as Bahdanua, scaled dot product (SDP), and Luong scaled dot product self-attention for spam email filtering. We evaluate our approach on various datasets, including Trec spam, Enron spam emails, SMS spam collections, and the Ling spam dataset, which constitutes a substantial custom dataset. All these datasets are publicly available. For the Enron dataset, we attain an accuracy of 99.97% using LSTM with SDP self-attention. Our custom dataset exhibits the highest accuracy of 99.01% when employing GRU with SDP self-attention. The SMS spam collection dataset yields a peak accuracy of 99.61% with LSTM and SDP attention. Using the GRU (Gated Recurrent Unit) alongside Luong and SDP (Structured Self-Attention) attention mechanisms, the peak accuracy of 99.89% in the Ling spam dataset. For the Trec spam dataset, the most accurate results are achieved using Luong attention LSTM, with an accuracy rate of 99.01%. Our performance analyses consistently indicate that employing the scaled dot product attention mechanism in conjunction with gated recurrent neural networks (GRU) delivers the most effective results. In summary, our research underscores the efficacy of employing advanced deep learning techniques and attention mechanisms for spam email filtering, with remarkable accuracy across multiple datasets. This approach presents a promising solution to the ever-growing problem of spam emails.
文摘With the rapid growth of the Internet in recent years, the ability to analyze and identify its users has become increasingly important. Authorship analysis provides a means to glean information about the author of a document originating from the internet or elsewhere, including but not limited to the author’s gender. There are well-known linguistic differences between the writing of men and women, and these differences can be effectively used to predict the gender of a document’s author. Capitalizing on these linguistic nuances, this study uses a set of stylometric features and a set of word count features to facilitate automatic gender discrimination on emails from the popular Enron email dataset. These features are used in conjunction with the Modified Balanced Winnow Neural Network proposed by Carvalho and Cohen, an improvement on the original Balanced Winnow created by Littlestone. Experiments with the Modified Balanced Winnow show that it is effectively able to discriminate gender using both stylometric and word count features, with the word count features providing superior results.
文摘In this paper, we propose a new online system that can quickly detect malicious spam emails and adapt to the changes in the email contents and the Uniform Resource Locator (URL) links leading to malicious websites by updating the system daily. We introduce an autonomous function for a server to generate training examples, in which double-bounce emails are automatically collected and their class labels are given by a crawler-type software to analyze the website maliciousness called SPIKE. In general, since spammers use botnets to spread numerous malicious emails within a short time, such distributed spam emails often have the same or similar contents. Therefore, it is not necessary for all spam emails to be learned. To adapt to new malicious campaigns quickly, only new types of spam emails should be selected for learning and this can be realized by introducing an active learning scheme into a classifier model. For this purpose, we adopt Resource Allocating Network with Locality Sensitive Hashing (RAN-LSH) as a classifier model with a data selection function. In RAN-LSH, the same or similar spam emails that have already been learned are quickly searched for a hash table in Locally Sensitive Hashing (LSH), in which the matched similar emails located in “well-learned” are discarded without being used as training data. To analyze email contents, we adopt the Bag of Words (BoW) approach and generate feature vectors whose attributes are transformed based on the normalized term frequency-inverse document frequency (TF-IDF). We use a data set of double-bounce spam emails collected at National Institute of Information and Communications Technology (NICT) in Japan from March 1st, 2013 until May 10th, 2013 to evaluate the performance of the proposed system. The results confirm that the proposed spam email detection system has capability of detecting with high detection rate.
文摘近年来,大型语言模型(large language model,LLM)技术兴起,类似ChatGPT这样的AI机器人,虽然其内部设置了大量的安全对抗机制,攻击者依然可以精心设计问答,绕过这些AI机器人的安全机制,在其帮助下自动化生产钓鱼邮件,进行网络攻击.这种情形下,如何鉴别AI生成的文本也成为一个热门的问题.为了开展LLM生成内容检测实验,从互联网某社交平台和ChatGPT收集了一定数量的问答数据样本,依据AI文本可获得条件的不同,研究提出了一系列检测策略,包含基于在线可获取AI对照样本的文本相似度分析、基于离线条件下使用统计差异性的文本数据挖掘分析、基于无法获得AI样本条件下的LLM生成方式对抗分析以及基于通过微调目标LLM模型本身构建分类器的AI模型分析,计算并比较了每种情况下分析引擎的检测能力.另一方面,从网络攻防的角度,针对检测策略的特点,给出了一些对抗AI文本检测引擎的免杀技巧.
