Online spam leads to malicious attacks such as hacking, phishing, and other problems that are more disruptive to Internet users. Online spam is currently known to be combated using detection and prevention techniques. Although these techniques can effectively curb these spreading messages/mails, it can be admitted that these attackers will persist in sending malicious links via unsolicited emails that can access and damage our system. Spammers easily construct phoney profiles and email addresses to target people unaware of these schemes. Spams always pose to be sent from a verified point by an authorized person; this makes it difficult to categorize as an attack or genuine mail. Consequently, it’s essential to detect fraudulent spam messages/mails. The paper discusses machine-learning techniques and the BERT model for text-based spam filter. The analysis was obtained over live data, extracted from a social media platform Reddit using PRAW. Comparison between different techniques was analysed to find the best fraudulent mail detection framework. It was observed that naive bayes had accuracy rate of 95% whereas accuracy of BERT framework came out to be 97% having rate of predicted spam was 86% and of ham was 99%