42/LOGOS-REPUTATION-INKINGUT

content/docs/rfcs/42/README.md

@@ -0,0 +1,177 @@
+---
+slug: 42
+title: 42/LOGOS-REPUTATION-INKINGUT
+name: Logos Node Reputation 
+status: raw
+category: Informational
+tags: logos/reputation
+editor: Mark Evenson <[email protected]>
+created: 04-AUG-2022
+revised: <2022-08-16 Tue 07:13>
+contributors:
+    - Álvaro Castro-Castilla
+---
+
+# Abstract
+
+We present Ikingut: an algorithm for purely local computation of peer
+node reputation based on their observed performance of a shared
+computation.
+
+# Background
+
+The following document describes the Ikingut reputation algorithm,
+designed to be used to augment the Logos consensus layer for various
+purposes.
+
+The use of objectives of defining a system for ongoing maintainance of
+node reputation is to
+
+   - Strengthen the Byzantine Fault Tolerance properties of the
+     consensus mechanism, possibly making the algorithm more resilient
+     to a higher number of byzantine nodes, or more aggressive
+     strategies.
+
+   - Providing a mechanism for bootstrapping networks, with a low
+     number of nodes.
+
+# Theory 
+
+## Ikingut Algorithm
+
+A deceptively simple algorithm for computing reputation of a other
+network nodes based on observed past performance.  
+
+### Concept
+
+The Ikingut algorithm for reputation is a local heuristic computed
+from a single source of information: *how many correct responses did a
+node give to queries*. The challenge in this metric is the definition
+of *correct response*. We define a correct response as the decision
+the node achieves on the bit value (by means of Glacier).
+
+This in essence summarizes the algorithm. There are implementation
+details and parameters that require some elaboration, as we will see
+below in further detail. Later on we will provide an experimental
+background to the specific design decisions.
+
+Other relevant characteristics of the algorithm are:
+
+- It’s simple. The complete implementation can be described in two
+  functions.
+- Lightweight. It doesn’t perform long computations nor use a lot of
+  memory.
+- Pluggable into Glacier [38/LOGOS-CONSENSUS-GLACIER][] and
+  Snowball. Inkingut is designed to fit well with swarm-like consensus
+  algorithms. Moreover, it augments the consensus algorithm without
+  altering it.
+- Adaptive. It processes network and behavior changes fast.
+- Dynamic. It introduces new nodes consistently and fairly.
+- It is robust against strategic adversaries. We will see these in more detail below.
+- No transitive trust. This avoids collusion, greatly simplifies. The
+  tradeoff is more time to build the reputation list.
+- Reasonable time to bootstrap reputation. With 10.000 nodes and k ~=
+  9 is in the order of 8-10 min.
+
+### Algorithm
+
+The Ikingut algorithm follows two parts, which run at every iteration
+of the consensus algorithm.
+
+1. On every agent vote:
+    1. Accumulate +1 for positive
+    2. Accumulate -1 for negative
+    3. Accumulate +1 in separate counter for *no response*
+2. On consensus decision finality. On every agent:
+    1. If the decision is NO, we need to invert the accumulated
+       reputation.
+    2. If the resulting reputation is positive, add it to the trust
+       score. A constant can be used for the reputation increasing
+       operation.
+    3. If the resulting reputation is negative, add it to the trust
+       score, using the following equation:
+       $$
+       t_i := min(at_i, t_i + br_i) \\
+       \text{where}\ 0 < a < 1 \ \text{and} \ b > 1 \\
+       $$
+       where $a$ and $b$ are constants used to manipulate how much
+       momentum is applied.
+
+
+**Pseudocode:**
+
+```python
+def ikingut_agent_trust(self, a, vote):
+		if vote == ConsensusAgent.YES:
+		    self.agent_votes_accum[a.unique_id] += 1
+    elif vote == ConsensusAgent.NO:
+        self.agent_votes_accum[a.unique_id] -= 1
+    else:
+        self.agent_votes_no_resp[a.unique_id] += 1
+
+def ikingut_model_trust(self):
+    # Compute new reputations
+    for (agent_id, rep_delta) in enumerate(self.agent_votes_accum):
+		    # If our decision is NO, we need to invert the accumulator
+		    if self.color == ConsensusAgent.NO:
+		        rep_delta = -rep_delta
+			  # No response is always negative score, so we substract it after
+		    # setting the right sign on the accumulator
+		    if rep_delta > 0:
+				    self.trust_scores[agent_id] += rep_delta
+		    elif rep_delta < 0:
+				    # Only apply momentum for punishment.
+				    self.trust_scores[agent_id] = min(self.trust_scores[agent_id] // 2,
+				                                      self.trust_scores[agent_id] + rep_delta)
+```
+
+**Min-multiplicative reputation punishment**
+
+The idea behind using a min function is that the punishment is much
+harsher if the current reputation of the node is high. The min
+function will select the highest reduction of reputation resulting
+from these two operations:
+
+- Multiplicative. This provides resiliency against patient attackers
+  that have built reputation long-term. Some back-of-the-envelope
+  calculations:
+    - In one year at 1000 tx/s (network size: 10000 nodes, k=9), the
+      expectation is around 0.9 req/s to each node. In 31536000
+      secs/year, the max positive score: 28,382,400 (all
+      reqs/node/year).
+    - With a 2x punishment: 25 reqs to lose of all reputation
+    - With a 3x punishment: 16 reqs to lose of all reputation
+
+    Note that these requests don’t need to be consecutive, just close
+    enough in time for the multiplicative punishment to take effect.
+
+- Linear. This ensures the following properties:
+    - Allow for new nodes to have a neutral score of zero.
+    - Bring the opportunity of building reputation faster.
+    - Ensure accumulation of negative scores for misbehaving nodes.
+
+# Implementation status
+
+Incomplete.  No publically known implementations.  
+
+# Sovereignty Considerations
+
+## Privacy
+
+Computation of reputation is purely local from observed behavior. 
+
+## Security
+
+No model constructed.  No current tests.
+
+# Copyright
+
+Copyright and related rights waived via
+[CC0](https://creativecommons.org/publicdomain/zero/1.0/).
+
+# Informative References
+
+1. [38/LOGOS-CONSENSUS-GLACIER](https://github.com/vacp2p/rfc/pull/512)
+
+2. [Rocket, Team, Maofan Yin, Kevin Sekniqi, Robbert van Renesse, and Emin Gün Sirer. “Scalable and Probabilistic Leaderless BFT Consensus through Metastability.” arXiv, August 24, 2020.](https://https://doi.org/10.48550/arXiv.1906.08936) 
+