This is one of my 2009 papers, presented by Randy Abrams and myself on behalf of ESETat the EICAR 2009 Conference in Berlin.
Anti-malware testing methodology remains a contentious area because many testers are insufficiently aware of the complexities of malware and anti-malware technology. This results in the frequent publication of comparative test results that are misleading and often totally invalid because they don’t accurately reflect the detection capability of the products under test. Because many tests are based purely on static testing, where products are tested by using them to scan presumed infected objects passively, those products that use more proactive techniques such as active heuristics, emulation and sandboxing are frequently disadvantaged in such tests, even assuming that sample sets are correctly validated.
Recent examples of misleading published statistical data include the ranking of anti-malware products according to reports returned by multi-scanner sample submission sites, even though the better examples of such sites are clear that this is not an appropriate use of their services, and the use of similar reports to generate other statistical data such as the assumed prevalence of specific malware. These problems, especially when combined with other testing problem areas such as accurate sample validation and classification, introduce major statistical anomalies.
In this paper, it is proposed to review the most common mainstream anti-malware detection techniques (search strings and simple signatures, generic signatures, passive heuristics, active heuristics and behaviour analysis) in the context of anti-malware testing for purposes of single product testing, comparative detection testing, and generation of prevalence and global detection data. Specifically, issues around static and dynamic testing will be examined. Issues with additional impact, such as sample classification and false positives, will be considered – not only false identification of innocent applications as malware, but also contentious classification issues such as (1) the trapping of samples, especially corrupted or truncated honeypot and honeynet samples intended maliciously but unable to pose a direct threat to target systems (2) use of such criteria as packing and obfuscation status as a primary heuristic for the identification of malware.