TCP insists on a handshake before any data flows. UDP does not bother. A UDP packet is simply sent, and if the receiving service wants to reply, it replies to whatever source address the packet claims to come from. That design is perfect for quick lookups such as DNS, for voice and video where a late packet is useless anyway, and for time synchronisation. It is also the root of the largest denial-of-service attacks ever recorded.
Reflection: borrowing someone else’s return address
Picture posting a letter to a large company asking for its full product catalogue, but writing your neighbour’s address as the sender. The catalogue goes to your neighbour. Do this with thousands of companies at once and your neighbour’s letterbox overflows.
Reflection attacks work the same way. Because UDP has no handshake, a server cannot easily verify that a request really came from the address in the packet. An attacker forges the victim’s IP address as the source and sends requests to many publicly reachable servers, called reflectors. The reflectors dutifully reply to the victim. The victim receives traffic from thousands of legitimate-looking servers, and the attacker’s own machines stay hidden behind them.
Amplification: when the reply is much bigger
Reflection becomes truly dangerous when the reply is larger than the request. That ratio is the amplification factor. Some protocols answer a small query with a very large response, for example DNS servers returning large record sets, old NTP servers that supported a diagnostic command listing recent clients, and exposed memcached servers. CISA’s published guidance on UDP-based amplification lists DNS at roughly 28 to 54 times and the old NTP diagnostic at up to about 556 times. Memcached was far worse; in 2018 it powered an attack on GitHub that peaked at about 1.35 Tbps.
A simple illustrative calculation shows the leverage. If an attacker controls machines able to send 100 Mbps of forged requests and the chosen protocol amplifies 50 times, the victim receives around 5 Gbps. Many Indian businesses connect their data centre or office through links of 1 Gbps or less, so the line saturates long before any server is stressed. No firewall rule on the victim’s side can help once the pipe itself is full; the packets must be dropped upstream.
Three shapes of DDoS
Amplification is one technique. It helps to sort denial-of-service attacks by what they exhaust, because each needs a different response.
| Shape | What runs out | Examples | Where mitigation happens |
|---|---|---|---|
| Volumetric | Network bandwidth | UDP reflection and amplification, large botnet floods | Upstream: ISP filtering, cloud scrubbing centres, anycast networks |
| Protocol (state exhaustion) | Connection tables in servers, firewalls, load balancers | SYN floods, fragmented packet floods, connection floods | SYN cookies, SYN proxies, firewall tuning, scrubbing |
| Application layer | CPU, memory, database capacity of the application | HTTP request floods, repeated expensive searches or logins | WAF, rate limiting, caching, bot detection, CAPTCHA challenges |
Volumetric attacks are measured in bits per second, protocol attacks in packets per second, and application attacks in requests per second. A low-volume application attack of a few thousand well-chosen requests per second can take down a site that would survive a 50 Gbps UDP flood behind a good scrubbing provider. Attackers frequently combine shapes, starting a volumetric flood to distract the team while a slower application attack runs in parallel.
Being a good neighbour: the reflector problem
There is an uncomfortable side to this topic. Every amplification attack needs reflectors, and reflectors are ordinary misconfigured servers belonging to someone else, perhaps your organisation. An open DNS resolver answering queries from the whole internet, an exposed memcached instance on a cloud VM, or an old NTP daemon with diagnostic features enabled can all be recruited without being “hacked” in any traditional sense.
- Resolvers should only answer queries from your own networks.
- Services such as memcached and Redis should never listen on public interfaces.
- Network providers should apply source address validation, described in the IETF best current practice BCP 38, so packets with forged source addresses never leave their network. If spoofing were impossible everywhere, reflection would not work.
CERT-In and other national CERTs periodically notify organisations about exposed services found in internet-wide scans, so ignoring an open resolver is also an operational and compliance risk.
Planning a defence before the day it happens
The worst time to buy DDoS protection is during an attack. A realistic plan for a business with public-facing services includes a few decisions made in advance.
First, know your capacity: the bandwidth of each internet link, the connection limits of your firewall and load balancer, and the request rate your application can sustain. Without these numbers, you cannot tell an attack from a busy sale day.
Second, decide where volumetric traffic will be absorbed. Options include an always-on CDN or cloud protection that fronts your website, an on-demand scrubbing service that your traffic is rerouted to during an attack, or DDoS protection offered by your ISP. On-demand services are cheaper but involve a switching delay, and they only help if your DNS or routing changes are rehearsed.
Third, harden the application. Cache what can be cached, put strict per-client limits on expensive operations such as search, OTP generation and login, and make sure the origin servers only accept traffic from the protection provider. A common mistake is putting a CDN in front of a website while the origin’s real IP address remains publicly discoverable, letting attackers bypass the shield entirely.
Finally, watch the right metrics. Bits per second catches volumetric floods. Packets per second and half-open connection counts catch protocol attacks. Requests per second per endpoint and the ratio of error responses catch application attacks. A dashboard showing only bandwidth will tell you everything is fine while the checkout page dies.
Written by the RP Strativa Team. Bandwidth and attack figures in the worked example are illustrative.
Leave a Reply