Bi-LSTM

Concept

Bidirectional Long Short-Term Memory (Bi-LSTM) is a recurrent neural network architecture that processes sequences in both forward and backward directions, enabling each output to encode information from both past and future context. It is widely used as a sequence encoder and classifier for tasks where local and bidirectional dependencies matter, including network-evasion detection, phishing URL classification, and CPU basic-block embedding for fuzzing.

First seen 7/10/2026

Last seen 7/10/2026

Evidence 1 chunks

Wiki v1

WIKI

Bi-LSTM

Bidirectional Long Short-Term Memory (Bi-LSTM) is a recurrent neural network architecture that runs two LSTM passes over the same sequence—one forward and one backward—and typically combines their hidden states (concatenation or summation) so that each position's representation is informed by both past and future tokens ^[1]. The bidirectionality makes Bi-LSTM particularly suitable for tasks in which the meaning of a token depends on context that lies on both sides of it in the sequence, rather than purely on preceding tokens.

Inductive bias and motivation

READ FULL ARTICLE →

NEIGHBORHOOD

No graph connections found for this entity yet. It may appear in future ingestion runs.

explore full graph →

RELATIONSHIPS

1 connections

Semantic-Aware Basic Block Encoder part of → 100% 1e

The Bi-LSTM is the backbone encoder of the Semantic-Aware Basic Block Encoder.

LINKED ENTITIES

1 links

Semantic-Aware Basic Block Encoder PART_OF Extracted graph relationship

CITATIONS

5 sources

5 citations — click to expand

[1] Bi-LSTM is adopted as the encoder backbone for the Semantic-Aware Basic Block Encoder because basic blocks are short, strictly ordered, single-entry/single-exit instruction sequences whose semantics are dominated by local data and control dependencies. HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing

[2] Transformer- and linear-attention backbones (GPT-2, GPT-2-small, RWKV-7) were benchmarked against Bi-LSTM; Bi-LSTM converges more stably on a moderate corpus under train-from-scratch. HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing

[3] The BB Encoder is trained in two stages: Stage 1 self-supervised MLM with 15% token-level masking and instruction-level masking on RISC-V assembly; Stage 2 supervised similarity learning using a custom BB-Sim metric derived from BinSequence-style weighted sequence alignment. HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing

[4] Bi-LSTM with a Softmax classifier is used for network-evasion detection, encoding both past and future traits of network flows and reaching 96.1% average accuracy. Network evasion detection with Bi-LSTM model

[5] Bi-LSTM classifies URLs into benign, phishing, defacement, and malware by capturing sequential and contextual information, reaching 97% accuracy on over 650,000 URLs. Phishing URL Detection using Bi-LSTM