DSS 2026 — Real-Time Threat Detection for DeFi with Stream Engines Jason W Source video: https://www.youtube.com/watch?v=-vVB06bYJ1o Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: 8ac5d2a12122de2c2e05bd3ca9107e22ca885cf6d9b27c269f7bfc203a08c753 [0:01] Hello, I am Jason, an independent [0:04] security researcher and deeply [0:06] passionate about Web three security and [0:09] build build the tools and a system to [0:12] contribute back to the community. [0:15] Uh today I will talk about real-time [0:18] threat detection for D5. [0:21] I'm focusing on two core goals. First, [0:24] uh tag threat as early as possible. [0:27] Um [0:29] blockchain [0:30] are almost always discovered after the [0:33] fact. Funds are already drained that [0:37] makes this a unique [0:39] hard problem compared to the Web two [0:42] security. [0:43] The core question I am trying to answer [0:46] is straightforward. Can we spot attacks [0:49] fast enough to create a window for [0:52] anti-money laundering [0:54] and ideally even interrupt the hike. [0:58] Second, build this real-time detection [1:01] system under tight resources constraint. [1:05] I run a R E T H execution node in [1:09] minimal mode to keep overhead no. [1:12] Let's walk through the system [1:14] architecture and the data pipeline. [1:17] I build a custom extension that collects [1:20] transaction traces and the event [1:23] logs. All raw [1:25] elementary flows into a adapter [1:28] component. [1:29] This adapter normalized on-chain events [1:32] and enrich it will enrich [1:35] smart contracts, core graphs and [1:38] attached and attached security labels. [1:41] In that inside this adapter a 10 [1:44] propagation algorithm track [1:46] mixer protocols such as Tornado Cash. [1:50] It marks address founded through Tornado [1:53] Cash and propagate that 10 label [1:55] downstream. [1:58] Normalized enriched the events then into [2:01] an event pool. [2:03] A stream matching engine consume this [2:05] upstream event. I predefine a set of [2:08] matching patterns and evaluate them [2:12] against the live event stream. When a [2:16] pattern matches [2:18] the event stream will fire an alert and [2:22] push it [2:23] into our console. [2:26] From there a language model and human [2:29] researchers will triage the alert and [2:32] conduct a deeper investigation. [2:36] Okay, let me show you an example [2:38] detection case [2:40] detection case [2:42] triggered inside the adapter built from [2:45] call frame trace data. [2:47] We reconstruct the cross side paths like [2:50] this. [2:51] And EOA interact with the same wallet [2:54] supplying [2:56] and supplying a malicious smart [2:58] contract. This smart contract was [3:01] deployed by the same EOA which is [3:06] already tiny and labeled as Tornado Cash [3:09] funded. [3:10] The execution contains nested dedicate [3:14] calls. [3:15] Immediately this [3:17] immediately [3:19] automatically this replace the [3:21] underlying implementation of the safe [3:24] wallet and transfer full control to the [3:27] attacker's smart contract. [3:30] Then the stream matching engine it [3:31] supports several pattern types equally [3:34] including sequence patterns and absence [3:37] patterns. [3:38] Here is an example based on the layer [3:41] zero exploit pattern. [3:45] In this attack these two events have a [3:48] gap of two blocks. [3:50] My rule works like this. Once [3:53] once we observe the package verified we [3:56] will look backward to check whether a [3:59] valid corresponding source [4:02] package sending exist. [4:04] If no matching source event can be [4:06] found, the rule will trigger an alert. [4:11] And moving on the investigation demo, [4:14] after the stream matching engine push [4:18] an alert to the console, what we can do? [4:23] Okay, this is our console. You can see [4:25] all the alerts has been listed here. And [4:28] this is our data plane. You can click [4:31] and see the details. And if you want to [4:34] the [4:35] language model agent to do the [4:37] investigation, just click this and it [4:40] will give you a [4:42] analysis here. [4:43] And also let me have a look at the [4:47] control plane. [4:49] We all the all the predefined patterns [4:53] listed here.